{ "schema_version": 1, "reviewed_at": "2026-10-12", "assignment": ".cache/oct12-scan/fulltext/foundation-model-assignment.json", "assignment_sha256": "9ba93117ba8a6f8d708b9b39ed424d8dee24e67fff5e17daabc2f62138e5f87d", "scope": "bounded foundation-model review; explicit non-exhaustive handoff", "counts": { "assigned": 258, "owned_assignment_range": [ 0, 138 ], "owned_assigned": 139, "reviewed_total": 143, "reviewed_in_owned_range": 139, "prior_reviews_outside_owned_range": 4, "reviewed_main_method_and_experiments": 13, "reviewed_abstract_and_selected_method_evaluation": 130, "accepted_in_frozen_batch": 3, "reviewed_deferred_next_batch": 129, "rejected_or_out_of_scope": 11, "not_reviewed": 115, "owned_not_reviewed": 0, "downloaded_with_sections": 252, "downloaded_without_sections": 6, "official_pdf_fallbacks_reviewed": 4 }, "coverage_note": "负责输入下标0..138共139篇，目前已审139篇、0篇未审；另保留分工调整前4篇尾部review供另一代理复用。冻结范围外候选只记下一批且未实现；新增审查明确记录method/evaluation阅读字符范围，不宣称完整主章节/appendix或穷尽生产证据。", "google_meta_priority_note": "输入priority_review_required全部false；按原文author affiliation核验找到7个Google/Meta作者或合作者（前半3、先前已审尾部4），优先实际审查，非正文citation关键字自动纳入。", "reviews": [ { "arxiv_id": "2610.12602", "title": "Self-Preservation in Language Models", "published": "2026-10-08T18:24:02Z", "authors": [ "Dongkyu Cho", "Rumi Chunara", "Ali Behrouz" ], "affiliation": [ "Dongkyu Cho Rumi Chunara Ali Behrouz Affiliation: [ Affiliation: [ Affiliation: [ Email: dongkyu.cho@nyu.edu Email: rumi.chunara@nyu.edu Email: alibehrouz@google.com" ], "source_url": "https://arxiv.org/html/2610.12602v1", "source_sha256": "2e2ca79ebe6e8e0b0cbbcecd8fd2a8113dcbdfceaf604845e9b151371c08badf", "source_bytes": 361853, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Google coauthor verified from author email", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "Task Bank中的selector、frozen pre-update replay generator、learner分别执行；只混合incoming与自生成replay，held-out任务描述/答案/分数不进入选bank。", "evaluation_evidence": "§3含Qwen0.6/1.7/4B多seed公开任务、matched pre/post retention；4B incoming-aware selection优于random，但小模型无对应收益。", "caveats": [ "fixed example allocation以replay替换incoming有学习/保留trade-off，不能归因replay内容本身", "post-hoc观察SFT损伤的selector是analysis reference，不能部署成无gold的agent", "作者GitHub地址可访问但当前repository empty" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/umamicode/self-preservation", "status": "empty repository verified 2026-10-12" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12602v1#S2", "https://arxiv.org/html/2610.12602v1#S3" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12602.html", "source_is_research_input_not_authority": true, "assignment_index": 84 }, { "arxiv_id": "2610.11361", "title": "SepGen: Multi-Stem Audio-Video Separation and Generation in a Single Model", "published": "2026-10-08T06:51:59Z", "authors": [ "Aviad Dahan", "Rajaei Khatib", "Yonatan Bitton", "Idan Szpektor", "Lior Wolf", "Raja Giryes" ], "affiliation": [ "Aviad Dahan Affiliation: Tel Aviv University Rajaei Khatib Affiliation: Tel Aviv University Yonatan Bitton Affiliation: Google Idan Szpektor Affiliation: Google Lior Wolf Affiliation: Tel Aviv University Raja Giryes Affiliation: Tel Aviv University" ], "source_url": "https://arxiv.org/html/2610.11361v1", "source_sha256": "601890f045905d7abd7d526aa1074d34e8411720df458cfb737052967baadde5", "source_bytes": 586981, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Google authors verified", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "LTX2.5 audio stream增加两stem，mix/video不得attend stems且LoRA只在stem tokens启用；同timeline，caption block routing，separation→generation两阶段noise schedule与estimated-clean-mix inference。", "evaluation_evidence": "§4涵盖194生成场景、Veo/LTX外部场景、172真实录音、三/五seed、swap margins与听觉人类对比。", "caveats": [ "human A/B listening study不是生产online A/B", "blank/unrelated/no video不显著改变separation自动指标，不能夸大video依赖", "公开项目页抓取失败，paper声称code/checkpoints/datasets available尚需验证资产可下载", "不能用独立speaker class scorer或仅音频pipeline替代joint masked model" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://sepgen.github.io/", "status": "advertised in paper; fetch error, unresolved asset availability" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11361v1#S3", "https://arxiv.org/html/2610.11361v1#S4", "https://arxiv.org/html/2610.11361v1#S4.SS4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11361.html", "source_is_research_input_not_authority": true, "assignment_index": 136 }, { "arxiv_id": "2610.11316", "title": "MetaEncoder: Exploring the Limit of Bi-Encoders for Multimodal System One Decision Making with Natural Language Interface", "published": "2026-10-08T06:22:48Z", "authors": [ "Jianpeng Cheng", "Guangyu Sun", "Aashu Singh", "Benyu Zhang", "Haixing Dai", "Hossein Mansour", "Jiangfan Zhang", "Shlok Kumar Mishra", "Wei Sun", "Xuanming Cui", "Yanli Liu", "Qi Guo", "Max Xiangjun Fan", "Jun Xiao" ], "affiliation": [ "Jianpeng Cheng Guangyu Sun Aashu Singh Benyu Zhang Haixing Dai Hossein Mansour Jiangfan Zhang Shlok Kumar Mishra Wei Sun Xuanming Cui Yanli Liu Qi Guo Max Xiangjun Fan Jun Xiao Affiliation: Meta AI" ], "source_url": "https://arxiv.org/html/2610.11316v1", "source_sha256": "a838ee38f50a178939e66545558bd0c8eb5298604f13f3cd7a146207b972c608", "source_bytes": 117321, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Meta AI authors verified", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "Muse-Glimmer30B causal last-token单位范数bi-encoder，request/candidate共享网络；closed-set可将option list放request，open-set缓存独立candidate；单向InfoNCE，重复positive不作negative。", "evaluation_evidence": "§4 11 suites/190 tasks；官方eval repo与30B模型card可访问，retrieval与decision按不同完整candidate sets评测。", "caveats": [ "official eval README披露Video-MMMU frames和prompt在test set sweep；这些结果含selection contamination", "相对OpenJev在text-only decision/knowledge部分更弱，MMEBvideo/image/visdoc也非全胜", "30B weights gated且80GB/GPU全copy；需要用户授权访问/实际资源而非用heuristic替代" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/facebook/meta-encoder-eval", "status": "accessible author evaluation code verified 2026-10-12" }, { "url": "https://huggingface.co/facebook/meta-encoder", "status": "accessible model card; weights gated" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11316v1#S3", "https://arxiv.org/html/2610.11316v1#S4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11316.html", "source_is_research_input_not_authority": true, "assignment_index": 139 }, { "arxiv_id": "2610.11019", "title": "Mid-Training Language Models on Raw Video", "published": "2026-10-08T00:03:39Z", "authors": [ "Jaedong Hwang", "Xiaoqian Shen", "Ernie Chang", "Changsheng Zhao", "Chong Zhou", "Saksham Suri", "Qi Qian", "Zechun Liu", "Lemeng Wu", "Qinsi Wang", "Raghuraman Krishnamoorthi", "Wei Wen" ], "affiliation": [ "Jaedong Hwang Affiliation: Meta AI Affiliation: Massachusetts Institute of Technology Work done at Meta Xiaoqian Shen Affiliation: Meta AI Affiliation: KAUST Work done at Meta Ernie Chang Affiliation: Meta AI Changsheng Zhao Affiliation: Meta AI Chong Zhou Affiliation: Meta AI Saksham Suri Affiliation: Meta AI Qi Qian Affiliation: Meta AI Zechun Liu Affiliation: Meta AI Lemeng Wu Affiliation: Meta AI Qinsi Wang Affiliation: Meta AI Affiliation: Duke University Work done at Meta Raghuraman Krishnamoorthi Affiliation: Meta AI Wei Wen Affiliation: Meta AI" ], "source_url": "https://arxiv.org/html/2610.11019v1", "source_sha256": "3e54ff157d3b394cea7b414fb23c3a6c6dbd939c3c7bdfb55d3e6c03d8031107", "source_bytes": 176779, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Meta AI authors verified", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "真实raw video frames经SigLIP/projector按grid+newline+time序列输入Qwen3；two-layer head以cosine distance预测next visual token，target stop-gradient；midtraining后两arm同一image-text SFT。", "evaluation_evidence": "§4 6.99M clips/64.4B visual tokens，128H100；4video/10image/14text，0.3/0.5/1epoch轨迹与caption-target对照。", "caveats": [ "0.3epoch后vision收益plateau", "text capability preserved仅相对no-midtraining SFT arm；二者都比原Qwen3在math/reasoning低，不能称语言绝对无损", "须保留实际连续视觉回归和后续matchedSFT两阶段，不能用图像类别或caption loss替代", "所见github lighteval是评价依赖，不是作者训练源码" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11019v1#S3", "https://arxiv.org/html/2610.11019v1#S4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11019.html", "source_is_research_input_not_authority": true, "assignment_index": 157 }, { "arxiv_id": "2610.09202", "title": "Few Bits, One Law: Toward W2A4KV2", "published": "2026-10-06T22:55:36Z", "authors": [ "Kai Yi", "Tarek Elgamal", "Sruthikesh Surineni", "Vignesh Vivekraja", "Soumyadeep Ghosh", "Steven Li" ], "affiliation": [ "Kai Yi Affiliation: Meta AI Tarek Elgamal Affiliation: Meta AI Sruthikesh Surineni Affiliation: Meta AI Vignesh Vivekraja Affiliation: Meta AI Soumyadeep Ghosh Affiliation: Meta AI Steven Li Affiliation: Meta AI" ], "source_url": "https://arxiv.org/html/2610.09202v1", "source_sha256": "4c89751e111611a7e0ec5e9fdbcedd29d24dce7a1855fbdb70a29141324fc446", "source_bytes": 771395, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Meta AI authors verified", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "固定HDP rotation+L2 energy canonicalization，离线Gaussian scalar/vector frozen codebooks，联合W/A/K/V真实QAT；normalization-aware STE不是粗略identity整块gradient。", "evaluation_evidence": "§5跨LLaMA1/3/8B、Qwen1.7B，WT2/eight zero-shot tasks、8K continuation与generation；零QAT在W2A4的PPL极差，需真实训练。", "caveats": [ "每configuration单training seed，不能称稳定formal improvement", "§5.6明示GSM8K用于learning-rate selection，其score不可当isolated test", "canonical/control量化active levels不同，table为完整系统对照非仅rotation效应", "未找到所读paper中的作者code链接" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.09202v1#S3", "https://arxiv.org/html/2610.09202v1#S4", "https://arxiv.org/html/2610.09202v1#S5" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.09202.html", "source_is_research_input_not_authority": true, "assignment_index": 234 }, { "arxiv_id": "2610.08917", "title": "CARE: Certifying Acceleration for Vision-Language-Action Inference", "published": "2026-10-06T18:00:05Z", "authors": [ "Rui Liu", "Tong Zheng", "Jindong Gu", "Zhipeng Wang" ], "affiliation": [ "Rui Liu Affiliation: University of Maryland, College Park Tong Zheng Affiliation: University of Maryland, College Park Jindong Gu Affiliation: University of Oxford Zhipeng Wang Affiliation: Google" ], "source_url": "https://arxiv.org/html/2610.08917v1", "source_sha256": "304d8813c0603a87a2faba6aed20023384930934c40a8132e0d770fe2d627291", "source_bytes": 435988, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Google coauthor verified", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "paired episode AIF=reference success且candidate fail；独立profiling排序、task-balanced round、family correction的anytime e-process、failure-triggered reference cache、无certificate回退reference。", "evaluation_evidence": "§4 LIBERO四suite与Crafter、独立calibration/test、上万resampling；包括24候选、flowsteps、不同模型的closed-loop完整rollouts。", "caveats": [ "LIBERO/Crafter是实验deployment，不是生产traffic online A/B", "单步local差异不等价于whole-episode AIF", "低riskbudget安全来自fallback，不能删fallback来制造speedup", "需要真实paired episode outcome与测量latency；随机loss/手造数组只能mechanism diagnostic" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.08917v1#S3", "https://arxiv.org/html/2610.08917v1#S4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.08917.html", "source_is_research_input_not_authority": true, "assignment_index": 246 }, { "arxiv_id": "2610.13858", "title": "LapMerge: Differentiable Exact-Budget Token Merging in Linear Time", "published": "2026-10-09T17:07:17Z", "authors": [ "Magdalena Trędowicz", "Aleksandra Nowak", "Tomasz Steifer", "Łukasz Struski", "Szymon Płotka", "Jacek Tabor" ], "affiliation": [ "Magdalena Trędowicz  Aleksandra Nowak  Tomasz Steifer Łukasz Struski  Szymon Płotka  Jacek Tabor Jagiellonian University in Kraków Institute of Fundamental Technological Research, Polish Academy of Sciences Centre for Credible AI, Warsaw University of Technology magdalena.tredowicz@doctoral.uj.edu.pl" ], "source_url": "https://arxiv.org/html/2610.13858v1", "source_sha256": "0ef73f65845cdbc22fd590c855c5b166de110953ac63ff903518c38cc8d6b778", "source_bytes": 1080622, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "non-Google/Meta practical architecture", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "Laplace mixture CDF把ordered tokens按equal mass分成exact N intervals；prefix/suffix闭式barrier+feature aggregation和backward，不构造M×N；log-space避免指数overflow。", "evaluation_evidence": "§4 GH200大序列runtime、patient-level6seedTCGA、Qwen2.5VL POPE frozen replacement、3seedLRA。", "caveats": [ "POPE tuned-width在evaluation set选，只能ablation ceiling，不能正式比较", "30%budget gainCI跨零，Pathfinder均chance；TCGA不是第一名", "operator O(M+N)不自动包括soft-top-k sorting/router成本", "supplementary code package被paper声称存在，但所读links只有UNI2 encoder，作者资产尚待确认" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13858v1#S3", "https://arxiv.org/html/2610.13858v1#S4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13858.html", "source_is_research_input_not_authority": true, "assignment_index": 21 }, { "arxiv_id": "2610.13605", "title": "zip2zip++: Towards Inference-Time Token Compression with Minimal Quality Loss", "published": "2026-10-09T13:52:54Z", "authors": [ "Saibo Geng", "Andrea Gentilini", "Xinxian Ma", "Nathan Ranchin", "Yunzhen Yao", "Robert West" ], "affiliation": [ "Saibo Geng † † thanks: Equal contribution. Author order was determined by a dice roll. Affiliation: EPFL Email: saibo.geng@epfl.ch Andrea Gentilini 1 1 footnotemark: 1 Affiliation: Affiliation: Technical University of Munich Affiliation: Zuse School ELIZA Email: xinxian.ma@epfl.ch Xinxian Ma 1 1 footnotemark: 1 Affiliation: EPFL Email: yunzhen.yao@epfl.ch Nathan Ranchin Affiliation: ETH Zürich Email: robert.west@epfl.ch Yunzhen Yao Affiliation: EPFL Email: a.gentilini@tum.de Robert West Affiliation: EPFL Email: nathan.ranchin@ethz.ch" ], "source_url": "https://arxiv.org/html/2610.13605v1", "source_sha256": "daf1a80ee646a2b6d7b277efb16fa77137677d355e518bacc05bf7381a4717cc", "source_bytes": 478896, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "non-Google/Meta practical architecture", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "LZW动态hypertokens，untied in/out hyperencoders、identity residual初始化、size-aware RoPE、protected digit/special tokens、fault-tolerant guarded insertion及exact多segmentation概率。", "evaluation_evidence": "§5四公开instruction模型与matchedLoRA预算，多corpus byte/word likelihood、math/code/IFEval/RULER与cumulative ablations；作者训练代码repo可访问。", "caveats": [ "multi-view scoring收益与modeling改善必须分开", "质量差距约半闭合而未关闭；14B HumanEval/IFEval下降", "200B compressed additionalpretraining不compute-matched且GSM8K下降", "公开runtime用hf revision，不可把main raw training checkpoint当HF可直接加载格式" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/epfl-dlab/zip2zip-plus-plus", "status": "accessible author training/evaluation code verified 2026-10-12" }, { "url": "https://github.com/epfl-dlab/zip2zip-compression", "status": "author codec link in paper; not separately downloaded" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13605v1#S3", "https://arxiv.org/html/2610.13605v1#S4", "https://arxiv.org/html/2610.13605v1#S5" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13605.html", "source_is_research_input_not_authority": true, "assignment_index": 36 }, { "arxiv_id": "2610.13041", "title": "RAMP: Recurrent Adaptive Mixed Precision for Efficient Prefill in Recurrent-Depth Language Models", "published": "2026-10-09T05:57:33Z", "authors": [ "Ke Wan", "Chen Chen" ], "affiliation": [ "Ke Wan Affiliation: Department of Computer Science, University of Virginia Email: tbn5pj@virginia.edu Chen Chen Affiliation: Department of Computer Science, University of Virginia Email: zrh6du@virginia.edu" ], "source_url": "https://arxiv.org/html/2610.13041v1", "source_sha256": "1eba41d730c3d49c74e8ee6eecde3c93cbb663950961faa547faafa9222d97cf", "source_bytes": 264035, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "non-Google/Meta practical efficiency", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "同一shared recurrent block前K步W+A低precision、terminal BF16 cleanup；fixedT/matched precisionbudget逆序和随机schedule对照，理论依赖local contraction。", "evaluation_evidence": "§6 Huginn0125 T32/K24，frozen nested-context WT2/PG19 identicaltail64 targets；H100 nativeFP8 first-prefill同optimizedemptycache路径。", "caveats": [ "128/256context比BF16慢，只有long-context约1.16x", "PG19 NLL对逆序差异未统计resolve，representation排序强不等于所有metric显著", "nativeFP8 H100不能在A100用fake quant冒充speedup", "作者源码链接未在所读paper中找到" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13041v1#S2", "https://arxiv.org/html/2610.13041v1#S3", "https://arxiv.org/html/2610.13041v1#S4", "https://arxiv.org/html/2610.13041v1#S6" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13041.html", "source_is_research_input_not_authority": true, "assignment_index": 68 }, { "arxiv_id": "2610.13708", "title": "Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy", "published": "2026-10-09T15:02:35Z", "authors": [ "Naji Alazhar", "Wassel Alazhar" ], "affiliation": [ "Umans AI (原文作者区 p2.2 人工核验)" ], "source_url": "https://arxiv.org/html/2610.13708v1", "source_sha256": "ca614cd43e9e1c8ceb8ec50c98be73f552a35460a8011178cf61c66ebd28d1fc", "source_bytes": 308821, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "non-Google/Meta practical serving", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "requiredhotcache/tail与reclaimable optionalhistory分开；ready publication+host backup invariant；exact多querymiss dedup、不合并attention selections，draftwindowring与layergroupverificationstaging复用，acceptedprefix只commit已接受KV。", "evaluation_evidence": "§4 8×B300，GLM5.3 100K与DeepSeekV4.1Flash500K concurrency serving，同model kernels，loadstep/width/hotcache sweeps与真实payloadcopy profiles。", "caveats": [ "闭环concurrency benchmark不是生产online A/B", "GLMC256 speculation更少slots导致TTFT更高且output不优于non-specadaptive；低load不优于residentserving", "headlineoutput只按decodeGPU计，不包含prefillGPU", "未确认作者实现链接；必须先验证exact runtime和真实I/O，不能用抽象KV模拟器声称系统复现" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13708v1#S3", "https://arxiv.org/html/2610.13708v1#S4", "https://arxiv.org/html/2610.13708v1#S4.SS1" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13708.html", "source_is_research_input_not_authority": true, "assignment_index": 30 }, { "arxiv_id": "2610.13735", "title": "Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers", "published": "2026-10-09T15:30:08Z", "authors": [ "Yuhong Chou", "Zehao Liu", "Yuqi Pan", "Qian Liu", "Xianwei Chen", "Zhongzhu Zhou", "Jibin Wu" ], "affiliation": [ "Yuhong Chou † , Zehao Liu † ,\nYuqi Pan ‡ , Qian Liu § ,\nXianwei Chen † , Zhongzhu Zhou ⋄ ,\nJibin Wu † † The Hong Kong Polytechnic University ‡ University of Chinese Academy of Sciences § Independent Researcher ⋄ University of Sydney Corresponding author: Yuhong Chou: yuhong.chou@connect.polyu.hk; Corresponding author: Jibin Wu" ], "source_url": "https://arxiv.org/html/2610.13735v1", "source_sha256": "b7e186007a39f796527eb4a0ae02e1cabf842b6aa55f44f3a65de4cc439a41ae", "source_bytes": 303285, "status": "accepted", "qualifies_for_reproduction_consideration": true, "selection_priority": "frozen Oct12 batch", "in_frozen_16_paper_batch": true, "next_batch": false, "implementation_status": "core_mechanism_code_present_gpu_pending", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "fixed-windowbaseSWAprefill、outputtokenfullcausalcrossattention、同层完整KV与真实conversion/training。", "evaluation_evidence": "已实现并有CPU formula/cache训练smoke，GPU未完成；原文H100/longcontext结果与本地L1分开。", "caveats": [ "不可把固定窗口线性prefill说成kernelizedlinearattention", "原文部署推理不是生产online A/B" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13735v1#S3", "https://arxiv.org/html/2610.13735v1#S4" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13735.html", "source_is_research_input_not_authority": true, "assignment_index": 29 }, { "arxiv_id": "2610.13767", "title": "DualQuant: Dual-scale Calibration-free Weight-and-Activation Quantization for LLMs", "published": "2026-10-09T15:51:38Z", "authors": [ "Mahsa Salmani", "Sayeh Sharify", "Hesham Mostafa" ], "affiliation": [ "Mahsa Salmani Sayeh Sharify & Hesham Mostafa Affiliation: d-Matrix" ], "source_url": "https://arxiv.org/html/2610.13767v1", "source_sha256": "eb28173622f2e939a056f0fb7aa00e472f9682cf13bfad6c3e49b56a690ac8b4", "source_bytes": 676238, "status": "accepted", "qualifies_for_reproduction_consideration": true, "selection_priority": "frozen Oct12 batch", "in_frozen_16_paper_batch": true, "next_batch": false, "implementation_status": "core_mechanism_code_present_gpu_pending", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "black-boxquantizer、inner/outvariable splitting、闭式双inverse-scale投影坐标下降，weight-derivedactivationrescale再实际量化。", "evaluation_evidence": "已实现限定INT4/MXINT4software simulation，CPU公式单测通过；公开checkpoint GPU未完成。", "caveats": [ "本地默认部分linear、无packing/nativeFP4，不能转述RTX4090paper加速为A100measurement" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/MahsaMoonlike/dualquant", "status": "accessible author repo verified 2026-10-12" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13767v1#S4", "https://arxiv.org/html/2610.13767v1#S5" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13767.html", "source_is_research_input_not_authority": true, "assignment_index": 27 }, { "arxiv_id": "2610.12417", "title": "WOVEN: Weaving Visual World Modeling into Multimodal LLMs", "published": "2026-10-08T17:52:03Z", "authors": [ "Zheyu Fan", "Yue Zhang", "Mingkai Deng", "Kangrui Wang", "Qineng Wang", "Canyu Chen", "Jie Hao", "Xing Fan", "Chenlei Guo", "Eric P. Xing", "Mohit Bansal", "Manling Li" ], "affiliation": [ "Zheyu Fan 1 , Yue Zhang 3 , Mingkai Deng 2 , Kangrui Wang 1 ,\nQineng Wang 1 , Canyu Chen 1 , Jie Hao 4 , Xing Fan 4 , Chenlei Guo 4 ,\nEric P. Xing 2 , Mohit Bansal 3 , Manling Li 1 1 Northwestern University 2 Carnegie Mellon University 3 UNC Chapel Hill 4 Amazon Website Code Dataset" ], "source_url": "https://arxiv.org/html/2610.12417v1", "source_sha256": "d6a25232b04ca31e9ca7f1b2c30de215a8ed8e4317dabe1b7044130866676011", "source_bytes": 3149038, "status": "accepted", "qualifies_for_reproduction_consideration": true, "selection_priority": "frozen Oct12 batch omission", "in_frozen_16_paper_batch": true, "next_batch": false, "implementation_status": "core_mechanism_code_present_gpu_pending", "full_paper_read": false, "read_scope": "abstract/author information/main defining methods/main experiments and tables; production-evidence paragraph spot check; appendices not exhaustively read", "defining_method": "真实transition图像条件监督，source/action/reasoning三轴、真实VLMresponse-only SFT和optiontargetpermutation。", "evaluation_evidence": "已实现publicWOVENtrain mini-suite与真实VLM链路，source_id分组隔离；作者officialvalidation/test未发布，GPU未完成。", "caveats": [ "不能将本地train-grouped diagnostic称officialbenchmark", "当前无GRPO与26benchmark迁移" ], "online_ab": "no quantified production-traffic A/B found in reviewed original sections; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/mll-lab-nu/WOVEN", "status": "accessible author training code; evaluation code coming soon" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12417v1#S2", "https://arxiv.org/html/2610.12417v1#S3", "https://arxiv.org/html/2610.12417v1#A7.SS1" ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12417.html", "source_is_research_input_not_authority": true, "assignment_index": 90 }, { "arxiv_id": "2610.13663", "assignment_index": 0, "title": "A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models", "published": "2026-10-09T14:34:22Z", "authors": [ "Ivan Moloshnikov", "Dmitry Kunitsyn", "Alexander Sboev", "Maksim Skorokhodov", "Roman Rybka", "Yuanchao Liu" ], "affiliation": [ "Ivan Moloshnikov ivan-rus@yandex.ru Dmitry Kunitsyn Kunitsyn_DE@nrcki.ru Alexander Sboev sag111@mail.ru Maksim Skorokhodov Skorokhodov_MS@nrcki.ru Roman Rybka Rybka_RB@nrcki.ru Yuanchao Liu ycliuharbin@163.com organization=National Research Centre “Kurchatov Institute”,\naddressline=1 Akademika Kurchatova Square,\ncity=Moscow,\npostcode=123182,\nstate=Moscow,\ncountry=Russia organization=National Research Nuclear University MEPhI,\naddressline=31 Kashirskoye Shosse,\ncity=Moscow,\npostcode=115409,\nstate=Moscow,\ncountry=Russia organization=Moscow Institute of Physics and Technology,\naddressline=9 Institutskiy Lane,\ncity=Dolgoprudny,\npostcode=141701,\nstate=Moscow,\ncountry=Russia organization=Harbin Institute of Technology,\naddressline=No. 92 West Dazhi Street, Nangang District,\ncity=Harbin,\npostcode=150001,\nstate=Heilongjiang,\ncountry=China" ], "source_url": "https://arxiv.org/html/2610.13663v1", "source_sha256": "03f1c51af165fb192abb1f13eff3038977f22e81c4a9a8836183329e0d84eb20", "source_bytes": 798735, "source_json_sha256": "ce9a242647490e37dc2ee0ef5fd68bbae949dc21fb7d34887be66889bd4c2aa4", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "固定网络/训练流水线，只让 Analyst提出指定模块规格、Coder生成契约约束代码、Docker实测、Recorder入log；错误保留但不计成功预算，不能把整个模型搜索当同一机制。", "evaluation_evidence": "SNN MNIST/Digits与三PINN任务；正文区分best-run、物理守恒诊断与十次初始化重复。", "caveats": [ "部分设置single seed，API依赖与有限成功迭代预算", "低成本Random Choice和Qwen agent准确率接近；不能仅看headline最佳比值" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/sag111/llm-auto-tuning-results", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13663v1#S3", "https://arxiv.org/html/2610.13663v1#S4", "https://arxiv.org/html/2610.13663v1#S5", "https://arxiv.org/html/2610.13663v1#S8" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method: Multi-agent LLM framework", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 7073, 7773 ] ] }, { "section_index": 3, "heading": "4 Experimental setup", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 7, "heading": "8 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 2039, 2739 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13663.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13563", "assignment_index": 1, "title": "Not All Correct Fixes Run Alike: An Empirical Study of Patch Performance in Repository-Level Issue Resolution", "published": "2026-10-09T13:17:53Z", "authors": [ "Han Peng", "Xing Hu", "Feifei Niu", "Huan Li", "Xin Xia" ], "affiliation": [ "Han Peng Affiliation: The State Key Laboratory of Blockchain and Data Security, Zhejiang University , Hangzhou , China Affiliation: Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security , Hangzhou , China email: awdhanpeng@zju.edu.cn , Xing Hu Affiliation: The State Key Laboratory of Blockchain and Data Security, Zhejiang University , Hangzhou , China email: xinghu@zju.edu.cn , Feifei Niu Affiliation: Graz University of Technology , Graz , Austria email: feifeiniu96@gmail.com , Huan Li Affiliation: The State Key Laboratory of Blockchain and Data Security, Zhejiang University , Hangzhou , China email: huanlee@zju.edu.cn and Xin Xia Affiliation: The State Key Laboratory of Blockchain and Data Security, Zhejiang University , Hangzhou , China email: xin.xia@acm.org" ], "source_url": "https://arxiv.org/html/2610.13563v1", "source_sha256": "d22c56f27d08185f762cfefc035eb0bf42024a61a86a17be1b05f9e1ec4fdd36", "source_bytes": 412118, "source_json_sha256": "5a0a1634777c9664b0d8fc02877fb0d355e10b373a4f5771c6bbaa0d164513e8", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "将每个test-passing patch的官方F2P测试改为warmup+N次内部循环、G次独立进程；固定CPU/NUMA隔离，指令数和walltime以Welch ANOVA区分性能。", "evaluation_evidence": "SWE-bench verified与Live过滤可测、至少两种不同correct solution的262案例；developer gold仅评价参照，不能给修复agent。", "caveats": [ "仅可测correct-patch子集不是全体解决率", "F2P准备/断言成本污染和重复状态复用可能遮蔽真实修复路径" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/SWE-agent/mini-swe-agent", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/astropy/astropy/pull/7647", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/astropy/astropy/pull/7671", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13563v1#S2", "https://arxiv.org/html/2610.13563v1#S3", "https://arxiv.org/html/2610.13563v1#S4", "https://arxiv.org/html/2610.13563v1#S5" ], "evidence_sections": [ { "section_index": 1, "heading": "2. Preliminaries: From F2P Test to a Measurable Corpus", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 2, "heading": "3. Runtime Performance Evaluation with F2P Tests", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 11486, 12186 ] ] }, { "section_index": 3, "heading": "4. Results and Analysis", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 4, "heading": "5. Threats to Validity", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 735, 1435 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13563.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12874", "assignment_index": 2, "title": "Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory", "published": "2026-10-09T02:05:38Z", "authors": [ "Junyou Guo", "Chongyang Tao", "Yan He", "Yong Wang", "Shuai Ma" ], "affiliation": [ "Junyou Guo Affiliation: School of Computer Science and Engineering, Beihang University, Beijing, China Chongyang Tao Affiliation: School of Computer Science and Engineering, Beihang University, Beijing, China Affiliation: Project Lead.  Corresponding author. Yan He Affiliation: School of Computer Science and Engineering, Beihang University, Beijing, China Yong Wang Affiliation: Alibaba Group Shuai Ma Affiliation: School of Computer Science and Engineering, Beihang University, Beijing, China" ], "source_url": "https://arxiv.org/html/2610.12874v1", "source_sha256": "cc0ee64ed538c269d993d0b91cf56e97bbbb657ef48541df13375a87a7d5ab52", "source_bytes": 425539, "source_json_sha256": "4a399ba5252e6b9f33fe8e283d12525ba1d900cc68d957c82706b376592dd19b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "原聊天存储不改写；metadata-expanded BM25种子、DBSCAN时间局部证据发现、Update/Expand证据forest和保序linearization组合，回答模型不见reference。", "evaluation_evidence": "GroupMemBench/EverMemBench官方长程多人对话；topic/phase移除消融仍保留部分优势。", "caveats": [ "需复现时间聚类/关系结构，不能只做BM25代理", "没有已核验作者实现；metadata依赖与超参选择需隔离validation" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12874v1#S4", "https://arxiv.org/html/2610.12874v1#S5" ], "evidence_sections": [ { "section_index": 3, "heading": "4 T-Mem", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 7372, 8072 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 13473, 14173 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12874.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12668", "assignment_index": 3, "title": "Recurrent Harness for Agentic Continual Learning", "published": "2026-10-08T19:44:43Z", "authors": [ "Yuxin Xiao", "Zhiyun Lu", "Jianyu Wang", "Yangtian Zhang", "Yang Song", "Qifan Wang", "Bailin Wang" ], "affiliation": [ "Yuxin Xiao Affiliation: Massachusetts Institute of Technology Work done during an internship at Meta MRS Research. Zhiyun Lu Affiliation: Meta MRS Research Jianyu Wang Affiliation: Meta MRS Research Yangtian Zhang Affiliation: Yale University Yang Song Affiliation: Meta MRS Research Qifan Wang Affiliation: Meta MRS Research Bailin Wang Affiliation: Meta MRS Research" ], "source_url": "https://arxiv.org/html/2610.12668v1", "source_sha256": "777a79b0dfea8cf1a8e9b70b30e9342cc059205ef6426572dad193c1c5775502", "source_bytes": 469482, "source_json_sha256": "58c23d054ab3eb2119d5adde8d69b6352508dfba63a05e8977653696a16c8881", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "Meta MRS Research/MIT original author block", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "strict online forward-read检索过去反思，当前轨迹完成后逐个backward-revise被检索反思；修改只能源自原轨迹，当前轨迹仅反馈；revised reflection示范forward-write，无回放/validation搜索。", "evaluation_evidence": "CL-Bench、MemoryBench、MemoryArena三backbone；source-grounding/泄漏审计、固定k=5后post-hoc敏感性、token和API成本。", "caveats": [ "grounding并非100%，要保留未支持/泄漏比例", "总token低于Top-K Raw但Claude缓存后费用更高；不是online生产流量A/B" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12668v1#S3", "https://arxiv.org/html/2610.12668v1#S4", "https://arxiv.org/html/2610.12668v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Framework", "read_character_ranges_in_cached_section": [ [ 0, 9299 ] ] }, { "section_index": 3, "heading": "4 Experiment", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 7933, 8633 ] ] }, { "section_index": 4, "heading": "5 Analysis", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 12508, 13208 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12668.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11920", "assignment_index": 4, "title": "Event-Centric Memory with Query-Aware Graph Augmentation for Long-Term Conversational Agents", "published": "2026-10-08T13:17:59Z", "authors": [ "Yichen Liu", "Chunfeng Yuan", "Haowei Liu", "Wenjuan Li", "Zefeng Lin", "Bing Li", "Xu Chen", "Weiming Hu" ], "affiliation": [ "Yichen Liu Chunfeng Yuan Haowei Liu Wenjuan Li Zefeng Lin Bing Li Xu Chen and Weiming Hu † † thanks: $ ∙ $ Y. Liu, C. Yuan, W. Li, Z. Lin, B. Li, and W. Hu are with the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China. † † thanks: $ ∙ $ Y. Liu, C. Yuan, B. Li, and W. Hu are also with the School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China. † † thanks: $ ∙ $ C. Yuan, W. Li, Z. Lin, B. Li, and W. Hu are also with the Beijing Key Laboratory of Super Intelligent Security of Multi-Modal Information, Beijing, China. † † thanks: $ ∙ $ H. Liu is with Alibaba Group, China. † † thanks: $ ∙ $ B. Li is also with PeopleAI Inc., Beijing 100190, China. † † thanks: $ ∙ $ X. Chen is with the Gaoling School of Artificial Intelligence, Renmin University of China, Beijing 100872, China. † † thanks: $ ∙ $ W. Hu is also with the School of Information Science and Technology, ShanghaiTech University, Shanghai 201210, China. † † thanks: $ ∙ $ Corresponding author: Chunfeng Yuan (e-mail: cfyuan@nlpr.ia.ac.cn)." ], "source_url": "https://arxiv.org/html/2610.11920v1", "source_sha256": "ae6134dcf1fac694d8d7fb80a7ab98bfb78bd1afe01f78d4467abc3724f6ea14", "source_bytes": 853252, "source_json_sha256": "492ead3d2923e1b5ec112bcbec36bc39a728dbac7e85f66281ad9e8237147215", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "事件原子记忆和动态trace保存来源/当前与历史状态；hybrid retrieval+query rerank激活局部graph，关系感知GNN mean pooling+projector成为真实graph token供LLM。", "evaluation_evidence": "LoCoMo、LongMemEval-S、MemoryAgentBench conflict、BEAM到10M histories；judge重复10次，并有retrieval和预算敏感性。", "caveats": [ "训练graph encoder/projector为核心，纯文字关系prompt是代理", "参数sweep需另行核独立validation，不能将论文最佳点当本地default证据" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11920v1#S3", "https://arxiv.org/html/2610.11920v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "III Method", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 3, "heading": "IV Experiments", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 13953, 14653 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11920.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11899", "assignment_index": 5, "title": "Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System", "published": "2026-10-08T13:05:33Z", "authors": [ "Irene Weber" ], "affiliation": [ "Irene Weber Affiliation: Faculty of Mechanical Engineering , University of Applied Sciences Kempten ,\nKempten, Germany Email: irene.weber@hs-kempten.de Affiliation:" ], "source_url": "https://arxiv.org/html/2610.11899v1", "source_sha256": "8b10da673f7004c723b180ef63227d6f4c8801980c11cfc648c9eeab874d6f42", "source_bytes": 171271, "source_json_sha256": "9d04bf39c6211a8ce628ee70b96487316325e654f2602d1c19cf90d84aa28549", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "描述22研究/厂商系统的调用控制、agent数、tool和人工介入的四维架构词汇；不是提出可执行的新能力算法。", "evaluation_evidence": "illustrative corpus与单作者架构编码，无模型受控能力训练/benchmark。", "caveats": [ "厂商文档和单作者coding可产生描述偏差", "保留为术语背景，不进入实现候选" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/features/copilot", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/features/copilot/ai-code-editor", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/modelcontextprotocol", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/blog/smolagents", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11899v1#S3", "https://arxiv.org/html/2610.11899v1#S4", "https://arxiv.org/html/2610.11899v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "III Agents and Tools", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 6679, 7379 ] ] }, { "section_index": 3, "heading": "IV Forms of LLM-Integrated Applications", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 4, "heading": "V Discussion", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 3743, 4443 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11899.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11794", "assignment_index": 6, "title": "Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks", "published": "2026-10-08T12:06:54Z", "authors": [ "Haoyu Zhao", "Zhengxu Yu", "Zhiyuan He", "Meng Fang", "Rasul Tutunov", "Haitham Bou-Ammar", "Weilin Luo", "Jun Wang" ], "affiliation": [ "Haoyu Zhao 1 Zhengxu Yu 2 Zhiyuan He 2 Meng Fang 3 Rasul Tutunov 2 Haitham Bou-Ammar 1,2 Weilin Luo 2 Jun Wang 1,† 1 University College London 2 Huawei Noah’s Ark Lab, UK 3 University of Liverpool" ], "source_url": "https://arxiv.org/html/2610.11794v1", "source_sha256": "456fd9e557b2e5aa08e6eb155effbe17b27af2b8c0d70fa8dd91afff3d66e4f8", "source_bytes": 513766, "source_json_sha256": "e8a003cbb2fe2e93b82093a959a10b934043f292ec4de2ba6ba71ce7c7525aa1", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "冻结LLM生成可修订natural-language rulebook并编译可执行world model；预测误差驱动修订，接受必须同时LLM faithful判定与cell-exact历史transition replay；population共享观察探索。", "evaluation_evidence": "25 public ARC-AGI-3 games全部实际交互动作计费，不看environment source；Pong三开局episode闭环。", "caveats": [ "公开games上的系统结果不是hidden benchmark或生产A/B", "不能用gold environment程序/转移表替代学习与编译；population部分单level不改进" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11794v1#S2", "https://arxiv.org/html/2610.11794v1#S4" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Model-Based Recursive Self-Improvement", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 8834, 9534 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11794.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11482", "assignment_index": 7, "title": "Evaluating Local Language Model Agents for Reproducible Data Engineering: An Empirical Software Engineering Study of Mobility Workflows", "published": "2026-10-08T08:26:46Z", "authors": [ "Jorge García-Carrasco", "Javier Sanchis", "Alejandro Reina-Reina", "Alejandro Maté", "Juan Trujillo" ], "affiliation": [ "Jorge García-Carrasco Javier Sanchis Alejandro Reina-Reina Alejandro Maté Juan Trujillo Lucentia Research, Department of Software and Computing Systems, University of Alicante, Ctra. de Sant Vicent del Raspeig s/n, 03690 Sant Vicent del Raspeig, Alicante, Spain * Corresponding author: jorge.g@ua.es javier.sanchis@ua.es , alejandro.reina@ua.es , amate@dlsi.ua.es , jtrujillog@dlsi.ua.es Preprint. Submitted to Information and Software Technology (Elsevier), manuscript\nnumber INFSOF-D-26-01448; currently under review. Replication package: https://doi.org/10.5281/zenodo.21397610 ." ], "source_url": "https://arxiv.org/html/2610.11482v1", "source_sha256": "951965a04e196a8fce7214e8671c335b2e67677f0571dee4ba8ea510a3194d23", "source_bytes": 226662, "source_json_sha256": "5fc866fdf1330c483703b401b379551f11dfb0f67eaeeba110b51de421b05ce7", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "15 mobility artifacts任务、bounded workspace，deterministic parsing/execution/schema/domain checker；one-shot与tools闭环配置对照及失败日志。", "evaluation_evidence": "10本地Qwen配置×2模式×15任务×5重复，共1500 attempt，RTX4090；Q4与部分BF16。", "caveats": [ "Wilson pooled intervals不是所有data-engineering任务的population CI", "尺度/架构/quantization未完全解耦，不能把workspace whole-condition差异归因单一tool" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/ggml-org/llama.cpp", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11482v1#S3", "https://arxiv.org/html/2610.11482v1#S4", "https://arxiv.org/html/2610.11482v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Benchmark design", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 13422, 14122 ] ] }, { "section_index": 3, "heading": "4 Experimental setup", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 7560, 8260 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 14024, 14724 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11482.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11447", "assignment_index": 8, "title": "Closed-loop evaluation of LLM agents for embedded software development", "published": "2026-10-08T08:02:10Z", "authors": [ "Jorge García-Carrasco", "Sergio García-Carrasco", "Alejandro Maté", "Juan Trujillo" ], "affiliation": [ "Jorge García-Carrasco Sergio García-Carrasco Alejandro Maté Juan Trujillo Lucentia Research, Department of Software and Computing Systems, University of Alicante,\nCtra. de San Vicente del Raspeig s/n, 03690 Sant Vicent del Raspeig, Spain Universitary Institute of Materials Technology (IUTM), Universitat Politècnica de València,\nPlaça Ferrándiz i Carbonell, 03801 Alcoi, Spain * Corresponding author: jorge.g@ua.es ORCID: 0000-0003-3174-083X (J. García-Carrasco), 0009-0001-9854-0309 (S. García-Carrasco),\n0000-0001-7770-3693 (A. Maté), 0000-0003-0139-6724 (J. Trujillo) Published in Journal of Systems Architecture 179 (2026) 103937, https://doi.org/10.1016/j.sysarc.2026.103937 . © 2026 The Authors.\nPublished by Elsevier B.V. under the CC BY 4.0 license\n( https://creativecommons.org/licenses/by/4.0/ )." ], "source_url": "https://arxiv.org/html/2610.11447v1", "source_sha256": "b3aec43cd6a8e63e02c3c258c9a13b3cdae27665629fdb6eef8fbb69a664a252", "source_bytes": 245971, "source_json_sha256": "4d8215d8c5906e73e090edbe68f7a0651b32f1575638196cdcedef0521058672", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "plain-text firmware工程handoff+QEMU ESP32 constrained workspace，在one-shot/self-verify/CI red-green/oracle四种可见反馈下同pi agent wrapper执行build/runtime修复。", "evaluation_evidence": "7模型×5任务×4反馈×3重复420runs，同RTX4090/llama.cpp/GGUF环境仅local latency可比。", "caveats": [ "真实能力以self-verify为主要设置；oracle详细hidden反馈必须单列diagnostic", "QEMU simulation不是物理device部署；fixed五任务不代表嵌入式全能力" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/badlogic/pi-mono/tree/main/packages/coding-agent", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/espressif/esp-idf", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/ggml-org/llama.cpp", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/jgcarrasco/closed_loop_evaluation_agents_embedded", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11447v1#S3", "https://arxiv.org/html/2610.11447v1#S4", "https://arxiv.org/html/2610.11447v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Benchmark Design", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 3, "heading": "4 Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 15300, 16000 ] ] }, { "section_index": 4, "heading": "5 Discussion", "read_character_ranges_in_cached_section": [ [ 0, 1100 ], [ 5613, 6313 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11447.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11314", "assignment_index": 9, "title": "From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue", "published": "2026-10-08T06:21:45Z", "authors": [ "Zirui Liao", "Zhengxian Wu", "Zhuohong Chen", "Yunyao Yu", "Xiaoyu Liu", "Yifan Xu", "Haoqian Wang" ], "affiliation": [ "Zirui Liao Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Email: liaozr24@gmail.com Zhengxian Wu Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Zhuohong Chen Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Yunyao Yu Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Xiaoyu Liu Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Yifan Xu Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China Haoqian Wang Affiliation: Tsinghua University Shenzhen International Graduate School, Shenzhen, China" ], "source_url": "https://arxiv.org/html/2610.11314v1", "source_sha256": "7b5396b30f5fc13726c35857d533598014cfb35a35fee84115c185efebf8063e", "source_bytes": 434016, "source_json_sha256": "f12bb8f84e6965f2ac58c59f58719c750c6e2bdc77f4e2e57758e3aacbf4b081", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "PEC²F graph将主观Claim与Fact/Event分离，保留source/target和SUPPORTED_BY；同source/predicate按时间reconcile，跨speaker不合并；四deterministic operators由LLM意图控制。", "evaluation_evidence": "LoCoMo与LongMemEval两英文backbone、semantic-collapse synthetic probe、50样本人偏好；offline consolidation成本/graph增长。", "caveats": [ "不能把Claim直接变成无来源事实", "50例人工审查与LLM judge不能证明多语/真实连续部署；周期offline consolidation有时延" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Silent-Rain02/CogMem", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11314v1#S3", "https://arxiv.org/html/2610.11314v1#S4", "https://arxiv.org/html/2610.11314v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15039, 15539 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 10301, 10801 ] ] }, { "section_index": 5, "heading": "6 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 1300, 1800 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11314.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11287", "assignment_index": 10, "title": "REMORY: Learning Residual Memory for Context Compaction", "published": "2026-10-08T05:50:50Z", "authors": [ "Hanchen Xia", "Baoyou Chen", "Yutang Ge", "Naihao Deng", "Senqiao Yang", "Zilong Dong", "Weihao Yuan", "Siyu Zhu" ], "affiliation": [ "Hanchen Xia Baoyou Chen Affiliation: Shanghai Academy of AI for Science Fudan University Yutang Ge Naihao Deng Affiliation: Shanghai Jiao Tong University University of Michigan Senqiao Yang Zilong Dong Affiliation: The Chinese University of Hong Kong Alibaba Group Weihao Yuan Siyu Zhu Email: {xiahanchen, chenbaoyou}@sais.org.cn * Corresponding author. Affiliation: Shanghai Academy of AI for Science Fudan University Affiliation: Nanjing University Shanghai Innovation Institute" ], "source_url": "https://arxiv.org/html/2610.11287v1", "source_sha256": "4ccf9c975837526fc3d5c91f84407b2baf52fe23aa5a98e90070604ecf8af5ac", "source_bytes": 112098, "source_json_sha256": "ee5aaf63335093ad8c112d17e1479e1e5e771ae192f5b5f53ef638bca8a86ea0", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "summary-conditioned neural memory产生bounded soft tokens附summary后；frozen full-history teacher与compacted student的complete-vocab reverse-KL，stageI reconstruction初始化、stageII residual objective，仅memory phi训练。", "evaluation_evidence": "SummHay全部92query、额外位置等预算summary++、多longhorizonagent任务；失败保留，gold insight/citations仅scoring。", "caveats": [ "必须真实soft embedding network，摘要加文本不是核心实现", "case-study修复不能证明信息来源一定来自softmemory，archive工具仍可普通访问" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/zapier/AutomationBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mlabonne/open-perfectblend", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/zai-org/GLM-5.3-Flash", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11287v1#S2", "https://arxiv.org/html/2610.11287v1#S3" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Residual Memory", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 3840, 4340 ] ] }, { "section_index": 2, "heading": "3 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 7476, 7976 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11287.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.10858", "assignment_index": 11, "title": "RFChipAgent: Multi-Agentic AI Flow for Analog/RF Chip Design", "published": "2026-10-07T20:02:33Z", "authors": [ "Awani Khodkumbhe", "Yunfei Feng", "Raj Rangarajan", "Kevin Wang", "Kamal Sahota" ], "affiliation": [ "Awani Khodkumbhe Affiliation: Qualcomm Technologies, Inc., UC Berkeley email: , Yunfei Feng Affiliation: Qualcomm Technologies, Inc. email: , Raj Rangarajan Affiliation: Qualcomm Technologies, Inc. email: , Kevin Wang Affiliation: Qualcomm Technologies, Inc. email: and Kamal Sahota Affiliation: Qualcomm Technologies, Inc. email:" ], "source_url": "https://arxiv.org/html/2610.10858v1", "source_sha256": "30b48143443f57b1b1af428da06da2eeda556007e7f17cda0fcb806e720ee3c1", "source_bytes": 110111, "source_json_sha256": "88c3e5dbab5e26e5e2fa9f5eef2f21e6ad966d866aba716f73bda64caf9d4aba", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Qualcomm RFIC superagent组织private document FAISS、topology和schematic/testbench agent、TPE+CMA-ES simulator loop与trust-scored simulationDB。", "evaluation_evidence": "GF22FDSOI 60GHz two-stage LNA一个family的设计campaign；成本/时长为系统run描述。", "caveats": [ "PDK/真实EDA simulation是必要外部依赖，未核公开可运行资产", "与手工设计努力比较并非随机受控production AB；不能用解析toy circuit替代" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.10858v1#S3", "https://arxiv.org/html/2610.10858v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3. Proposed RFChipAgent Flow", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] }, { "section_index": 3, "heading": "4. Experiments and Evaluations", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 2051, 2551 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.10858.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.09772", "assignment_index": 12, "title": "Decoupling Logic from Persona: Structural Immunity of Edge LLM Agents to Context Pollution", "published": "2026-10-07T09:53:28Z", "authors": [ "Masaaki Nakatsu", "Reno Wang" ], "affiliation": [ "Masaaki Nakatsu (AO, Inc. / OrbLabs AG) Reno Wang (AO,\nInc.)" ], "source_url": "https://arxiv.org/html/2610.09772v1", "source_sha256": "2664e6eff1883971bd197cf01d7f72d0989a3567bcdd6c2f34f5f5b0fae41798", "source_bytes": 572489, "source_json_sha256": "4755fadecb995733fee355bff9ade9e190037239bba5e021153c0b61cf778af9", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "同INT4 base的What logic LoRA只读core/vitals，生成Micro-State；How persona adapter读history与structuredstate，deterministic router/cache状态变化失效；两次decode不是简单promptwordfilter。", "evaluation_evidence": "M2、Llama8B/Gemma4B、两task两persona五seed pollution ablations；低pollution无逻辑优势，4B polluted-control崩溃。", "caveats": [ "How评估使用固定Micro-State，不能广告完整error传播端到端能力；gold/provided state路径须保留diagnostic边界", "20样本adapter、keywordrubric、单机少run系统延迟；同prompt同seed byte-invariance是结构性质非普遍鲁棒性" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/ml-explore/mlx", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.09772v1#S3", "https://arxiv.org/html/2610.09772v1#S4", "https://arxiv.org/html/2610.09772v1#S5", "https://arxiv.org/html/2610.09772v1#S8" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Architecture: AO-DA", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 3220, 3720 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 6410, 6910 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] }, { "section_index": 7, "heading": "8 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 1843, 2343 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.09772.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.10622", "assignment_index": 13, "title": "WorldBench: Evaluating LLMs on Three.js Voxel World Generation", "published": "2026-10-07T08:52:08Z", "authors": [ "Krish Bakshi" ], "affiliation": [ "Krish Bakshi Email: krishrajeshbakshi@gmail.com" ], "source_url": "https://arxiv.org/html/2610.10622v1", "source_sha256": "1265b42b975f16534213ffcbbd1ce4763a7fcb60c20fe2f12b76c0bc1eae99a5", "source_bytes": 384017, "source_json_sha256": "9e74717825054b77e21f12c4e2d2069be4f408489d56c705d719c651ced80a81", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Three.js world评审同时探索运行渲染/时钟/nav和读源code；codequote需exactsource，视觉可测性质pixel验证，两个证据halves独立计分。", "evaluation_evidence": "单固定prompt五frontierworlds，365items双通道差异及人工构造feature-removal mutation测试。", "caveats": [ "未执行code仍获半分，未量化cross-family judge bias", "一个syntheticworldprompt，无human一致性完整验证；不是通用3D生成benchmark" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/KrishBakshi/worldbench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.10622v1#S3", "https://arxiv.org/html/2610.10622v1#S4", "https://arxiv.org/html/2610.10622v1#S5", "https://arxiv.org/html/2610.10622v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 The benchmark", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 4615, 5115 ] ] }, { "section_index": 3, "heading": "4 The judge: explore and read, trust neither", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 10027, 10527 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 9155, 9655 ] ] }, { "section_index": 5, "heading": "6 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 2350, 2850 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.10622.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.09240", "assignment_index": 14, "title": "Adversarial Images Hijack Web Agents from Visual Grounding to Browser Execution", "published": "2026-10-07T00:05:53Z", "authors": [ "Wanjing Han", "Levi Taiji Li", "Mu Zhang", "Yue Jiang", "Guanhong Tao" ], "affiliation": [ "Wanjing Han Levi Taiji Li Mu Zhang Yue Jiang Guanhong Tao Affiliation: University of Utah Affiliation: {wanjing.han, levili, muzhang, yue.jiang, guanhong.tao}@utah.edu" ], "source_url": "https://arxiv.org/html/2610.09240v1", "source_sha256": "7f17e1a422f411e1899779c1ddefeac464a3f1e789b5a998deb4bc98e4a415c9", "source_bytes": 408940, "source_json_sha256": "3aaeea6994fd7626c15c4caa5306f6876d96a4a357b225075c2c912532e15bc5", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "white-box局部image扰动按role-slot抽象、page recomposition和action-postprocessing dataflow对齐，离线优化grounding-to-execution而非仅rawresponse。", "evaluation_evidence": "四agent配置/六VLM、13公共站点和sandbox多step任务，含defenses和request variation。", "caveats": [ "威胁模型需whitebox且只能image不可改DOM/code/userrequest", "未来安全复现只限自有sandbox，不触发实际购物/雇佣等第三方browser执行；截断主实验未穷尽" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://codeql.github.com/", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/MoonTea0416/WebMirage", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.09240v1#S3", "https://arxiv.org/html/2610.09240v1#S4", "https://arxiv.org/html/2610.09240v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Threat Model", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 2014, 2514 ] ] }, { "section_index": 3, "heading": "4 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] }, { "section_index": 4, "heading": "5 Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.09240.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.09079", "assignment_index": 15, "title": "Large-scale Repository Engineering via Agent-Native Reusable Code Primitives", "published": "2026-10-06T20:20:50Z", "authors": [ "Haibo Jin", "Peng Kuang", "Xucheng Yu", "Jerry Wang", "Dehao Wu", "Haohan Wang" ], "affiliation": [ "Haibo Jin 1 Peng Kuang 1 Xucheng Yu 1 Jerry Wang 1 Dehao Wu 1 Haohan Wang 1 1 University of Illinois Urbana-Champaign, USA † † thanks: Corresponding Author: haohanw@illinois.edu" ], "source_url": "https://arxiv.org/html/2610.09079v1", "source_sha256": "559a07bff49ae267844b7d1c56e4de5c78787c8e7e64dd29c6340b0c70b60d13", "source_bytes": 1136799, "source_json_sha256": "132ab76911075a2d138aa87d325384c273d9c6524dd0ae31e5de0ca66b38c1cd", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "CodeFace primitive携带interface/dependency/tests/provenance且residentLLM适配；LEGO做cross-componentconstraint resolution和真实executedtest诊断定点再激活，budget5/module。", "evaluation_evidence": "522 packages、13backbones、真实native suites空packagefloor到original-sourceceiling；有remine/disjoint控制与test-special-casing抽样审查。", "caveats": [ "公开native tests同时spec与grading，不能当独立hidden泛化test", "要核target code/primitive overlap和按task冻结环境，不能把retrievedcode直接copy当定义机制" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.09079v1#S3", "https://arxiv.org/html/2610.09079v1#S4", "https://arxiv.org/html/2610.09079v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 8899, 9399 ] ] }, { "section_index": 3, "heading": "4 LEGO-REPO Benchmark", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 2889, 3389 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.09079.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.08954", "assignment_index": 16, "title": "RACER: Reflective Agent Coupling Query Interpretation and Tool-Based Retrieval for Frame Selection in Long Video Understanding", "published": "2026-10-06T18:21:13Z", "authors": [ "Yiyang Huang", "Yitian Zhang", "Yizhou Wang", "Jianglin Lu", "Qihua Dong", "Hailing Wang", "Huimin Zeng", "Mingyuan Zhang", "Yun Fu" ], "affiliation": [ "Yiyang Huang Yitian Zhang Yizhou Wang Jianglin Lu Qihua Dong Affiliation: Department of Electrical and Computer Engineering, Northeastern University Email: huang.yiyan@northeastern.edu Hailing Wang Huimin Zeng Mingyuan Zhang Yun Fu Affiliation: Department of Electrical and Computer Engineering, Northeastern University Affiliation: Khoury College of Computer Science, Northeastern University Email: y.fu@northeastern.edu" ], "source_url": "https://arxiv.org/html/2610.08954v1", "source_sha256": "6563e3ee1075a0019203237bb8d0d87bf1dd4c0fa62c2289dd16582a24eacb33", "source_bytes": 278892, "source_json_sha256": "5fef231d3ae6e4525717046c5694114f7d3b4c24b2d6f27ea850cb65ec7f3879", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "轻量VidLLM把complexquery改为M=3 evidence-subqueries，CLIP定位真实video frames，反馈frames反思query再检索；query模型不直接决定frames。", "evaluation_evidence": "Video-MME无字幕、MLVU、LongVideoBench，四answeringVLM32-frame budget；3Bqueryinterpreter+ViT-L/14工具。", "caveats": [ "需真实video/frame embeddings+queryreflection，文本时间戳selector代理不成立", "未核作者repo和独立validation超参选择，production AB未建立" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.08954v1#S3", "https://arxiv.org/html/2610.08954v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 6638, 7138 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.08954.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13935", "assignment_index": 17, "title": "Hide and Seek: HIDE Concepts to SEEK explanations for black box models", "published": "2026-10-09T17:56:59Z", "authors": [ "Bhanu Tokas", "Rahul Nair", "Rockwell Jackson", "Hannah Kerner" ], "affiliation": [ "Bhanu Tokas Affiliation: Arizona State University Email: {btokas}@asu.edu Rahul Nair Affiliation: Arizona State University Email: {rnair21}@asu.edu Rockwell Jackson Affiliation: Arizona State University Email: {rpjacks2}@asu.edu Hannah Kerner Affiliation: Arizona State University Email: {hkerner}@asu.edu" ], "source_url": "https://arxiv.org/html/2610.13935v1", "source_sha256": "ae0e220f190e59178cc088d065e17bf1f7d76848d388c05aa1e30bad4e78c68e", "source_bytes": 425433, "source_json_sha256": "a7f1c063c0a8197fb04f81692ee896960ea15178f77dbf0e565cbabe3d9adc31", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "给定concept文本，经VLM检索concept-aligned真实图像并localize/mask对应区域；black-box model原图/遮挡图输出差构成global/local concept attribution，非梯度weight-access方法。", "evaluation_evidence": "CelebA/COCO-Stuff与三视觉classifier、captioning与crop segmentation应用；concept embedding preprocessing有不同VLM消融。", "caveats": [ "masking造成分布改变，行为差异不是无条件因果机制识别", "作者匿名supplement code声明未独立核可下载实现，需真实VLM localization不能类标签scorer代理" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13935v1#S3", "https://arxiv.org/html/2610.13935v1#S4", "https://arxiv.org/html/2610.13935v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 HIDE AND SEEK", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9519, 9919 ] ] }, { "section_index": 3, "heading": "4 Experiments and Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6543, 6943 ] ] }, { "section_index": 4, "heading": "5 Hide and Seek in Real World Applications", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6732, 7132 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13935.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13925", "assignment_index": 18, "title": "Can machines reason non-verbally?", "published": "2026-10-09T17:54:36Z", "authors": [ "Shaheer U. Saeed", "Zion Tse", "Gregory Slabaugh", "Daniel C. Alexander", "Yipeng Hu" ], "affiliation": [ "Shaheer U. Saeed 1,2,3,∗ , Zion Tse 1,2 , Gregory Slabaugh 1 , Daniel C. Alexander 3 , Yipeng Hu 3 1 Digital Environment Research Institute, Queen Mary University of London 2 Centre for Bioengineering, Queen Mary University of London 3 UCL Hawkes Institute, University College London ∗ Correspondence e-mail: shaheer.saeed@qmul.ac.uk" ], "source_url": "https://arxiv.org/html/2610.13925v1", "source_sha256": "4dcb4a6c0955369006da7e0d7c5166515b9db410a44ab916834d65d1e294f51f", "source_bytes": 213543, "source_json_sha256": "0a73fde88dedb695623ac0bbe887c7a4aa397a022da89434dc08162de476a449", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "GenVerOpt按sample生成多candidate、task-specific verifier评估并optimization/refinement，四task分别实例化，不能只复用抽象Generate-Verify口号。", "evaluation_evidence": "visual IQ和三medical image case studies与human/模型比较；本轮只审bounded recipe/主实验段，未完整核四个实例/validation。", "caveats": [ "human/clinical headline不代表通用机器reasoning或临床部署有效性", "下一批须逐个核verifier的训练/label隔离与实际generator/optimizer，不能gold评分器替代" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13925v1#Sx1", "https://arxiv.org/html/2610.13925v1#Sx2" ], "evidence_sections": [ { "section_index": 0, "heading": "A recipe for reasoning - Generate, Verify, Optimise", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3985, 4385 ] ] }, { "section_index": 1, "heading": "Machine visual reasoning in action", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13984, 14384 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13925.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13917", "assignment_index": 19, "title": "Designing Protein Structures from Natural Language", "published": "2026-10-09T17:51:55Z", "authors": [ "Joseph Lee", "Yanjing Li", "Bo Qiang", "Gregory Lau", "Allyson Ettinger", "Noah A. Smith", "Rory Donovan-Maiye", "David Baker", "Hanna Hajishirzi", "Pang Wei Koh" ], "affiliation": [ "Joseph Lee Affiliation: University of Washington Yanjing Li Affiliation: University of Washington Bo Qiang Affiliation: University of Washington Gregory Lau Affiliation: National University of Singapore Allyson Ettinger Affiliation: Allen Institute for AI (Ai2) Noah A. Smith Affiliation: University of Washington Affiliation: Allen Institute for AI (Ai2) Rory Donovan-Maiye Affiliation: Allen Institute for AI (Ai2) David Baker Affiliation: University of Washington Affiliation: Howard Hughes Medical Institute Hanna Hajishirzi Affiliation: University of Washington Pang Wei Koh Affiliation: University of Washington" ], "source_url": "https://arxiv.org/html/2610.13917v1", "source_sha256": "d6d74fb9decde0583f0addccedc7c60e2fdb131a9e82a72ac1e2a11f762afbf0", "source_bytes": 851422, "source_json_sha256": "18ce4121fdeb53149bf97acf34e2b03c00dcb80542182eff94012ccd258756e9", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Prompteina用1.8M structure/50M+captions对pretrained all-atom Proteina-Complexa text conditioning训练，residue-text对齐finetune；SDE+CFG生成backbone/sidechainlatents，frozenVAE解码结构/序列。", "evaluation_evidence": "structural properties、20binder targets和iterative in-silico prompt revision；GM2A实际89designs SPR assay发现9binders。", "caveats": [ "实验binding不代表临床/功能疗效；结构confidence只是in-silico代理", "大规模训练/data与专用结构模型依赖未独立核public可运行资源，不能LLM文本答结构替代" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/PaddlePaddle/PaddleHelix/tree/dev/apps/helixprotx", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/jleechung/prompteina", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13917v1#S2", "https://arxiv.org/html/2610.13917v1#S3", "https://arxiv.org/html/2610.13917v1#S5" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Text-guided Protein Structure Generation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6795, 7195 ] ] }, { "section_index": 2, "heading": "3 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "5 Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2321, 2721 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13917.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13910", "assignment_index": 20, "title": "When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors", "published": "2026-10-09T17:50:03Z", "authors": [ "Shubh Garg", "Chen Shani" ], "affiliation": [ "Shubh Garg Affiliation: Thapar Institute of Engineering and Technology Email: shubhgarg265@gmail.com Chen Shani Affiliation: Tel Aviv University Email: cshani@tauex.tau.ac.il" ], "source_url": "https://arxiv.org/html/2610.13910v1", "source_sha256": "9dd1425bbab2c1f11831e4eb49d40fb3f64fd822a6a3381d260221f72a2bfbf7", "source_bytes": 546462, "source_json_sha256": "f1e6de488e7c8ab382d966c623df61acf7bd052331d620823eb3726ccac5f050", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "CASP-VL在VisualGenome scenegraph只改一个concept/rendercaption；linearprobe仅原caption训练，EICS问perturbedcaption是否随改动稳定解码，BoW作为lexical baseline。", "evaluation_evidence": "跨CLIP/SigLIP/LLaVA layers、predicate-token与EOS/mean pooling、outputbenchmark对照及bootstrapCI。", "caveats": [ "representation诊断不是模型训练提升；probe target仅评估，不能输入被评模型答案", "部分spatialEICS置信区间到零；不能把pooling诊断当全部空间理解失败" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Shubhf/CASP-VL", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13910v1#S3", "https://arxiv.org/html/2610.13910v1#S4", "https://arxiv.org/html/2610.13910v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 CASP-VL and EICS", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5463, 5863 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14622, 15022 ] ] }, { "section_index": 4, "heading": "5 Analysis and Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3975, 4375 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13910.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13840", "assignment_index": 22, "title": "Studying Comics at Scale with Artificial Intelligence: The PAGE Framework and Three Studies of American Comics and Japanese Manga", "published": "2026-10-09T16:49:29Z", "authors": [ "Emilio Ferrara" ], "affiliation": [ "Emilio Ferrara Email: emiliofe@usc.edu Affiliation: Thomas Lord Department of Computer Science, University of Southern California, Los Angeles, CA, USA" ], "source_url": "https://arxiv.org/html/2610.13840v1", "source_sha256": "9f20dfdb7de9af74a03125fa07cef537f64700b8932c947f5debb84ac4535e9a", "source_bytes": 239571, "source_json_sha256": "61aba7172c5fa2007be9fc71bc0c41387976ee027734040274c02f701b9ded5b", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "PAGE作为comics研究instrument：parse panel/OCR/bibliography、固定codebook VLM annotation、独立human/reference grounding与genre/publisher controlled趋势估计。", "evaluation_evidence": "3929Americanbooks/87manga；name-withheld人标gender核验，修正图文页码偏移后retrieval比dense增益CI跨零。", "caveats": [ "主要贡献为人文档案实证instrument，不是本批foundation模型/能力算法adapter", "原错误页码join曾夸大image优势；数据license和duplicateissues需保留" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [], "evidence_sections": [ { "section_index": 2, "heading": "3 The PAGE framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9852, 10252 ] ] }, { "section_index": 3, "heading": "4 Data", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1812, 2212 ] ] }, { "section_index": 4, "heading": "5 Grounding the instrument before use", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4259, 4659 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13840.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13832", "assignment_index": 23, "title": "SpatialASM: Internalizing Active Mentalization for Spatial Reasoning", "published": "2026-10-09T16:39:50Z", "authors": [ "Yuhong Liu", "Shenxi Wu", "Junyi Chi", "Beichen Zhang", "Xuanlang Dai", "Jiaqi Wang", "Yuhang Zang", "Dahua Lin" ], "affiliation": [ "Yuhong Liu Affiliation: Shanghai AI Laboratory Shenxi Wu Affiliation: Shanghai AI Laboratory Junyi Chi Affiliation: Shanghai AI Laboratory Affiliation: Fudan University Beichen Zhang Affiliation: The Chinese University of Hong Kong Affiliation: Shanghai AI Laboratory Xuanlang Dai Affiliation: Shanghai AI Laboratory Affiliation: Fudan University Jiaqi Wang Yuhang Zang Dahua Lin Affiliation: Shanghai AI Laboratory Affiliation: CPII under InnoHK Affiliation: Shanghai Innovation Institute" ], "source_url": "https://arxiv.org/html/2610.13832v1", "source_sha256": "d95d3fc51bc40f356391bd959e8a39aca4082e0dbca7af461669ad357d3fbcb6", "source_bytes": 418339, "source_json_sha256": "813875d01cb98ba740befe4a2c685647a33ae7cd39da20c26ab8083160289929", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "训练时真实allocentric novel-view generator空间模拟；LVLM view-plan/NVG render后转caption的ASM轨迹SFT，自动perspective-taking tasks和GeoP几何一致性约束。", "evaluation_evidence": "四LVLM backbone、7spatial+3generalVQA；外置ASM与内部SFT/RL消融，单SFT/RL优势随task不同。", "caveats": [ "需真正NVG/render supervision，原问题追加spatialchainprompt不是复现", "novel-view generator及RL训练成本/验证切分待下一批更深核，不能claimonline AB" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/InternLM/SpatialASM", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13832v1#S3", "https://arxiv.org/html/2610.13832v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 SpatialASM", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14700, 15100 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12983, 13383 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13832.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13797", "assignment_index": 24, "title": "REFLECT: Role-Aware Detection of Sentiment, Emotion, Politeness, and Formality", "published": "2026-10-09T16:10:25Z", "authors": [ "Sangpil Youm", "Annabel Stocks Natalias", "Hayden Heninger", "Leah Jones", "Xiaoyang Jin", "Xiaoya Zhang", "Bonnie Dorr" ], "affiliation": [ "Sangpil Youm Annabel Stocks Natalias Hayden Heninger Leah Jones Xiaoyang Jin Xiaoya Zhang Bonnie Dorr" ], "source_url": "https://arxiv.org/html/2610.13797v1", "source_sha256": "cca38ceb36e59a86b83d93d11fece7fb136309daf76aa953f69781bb8e528e8c", "source_bytes": 186278, "source_json_sha256": "e82e00dbc218479024a4563a506c9d233beca25b0070a13bdb49f06e6b32aed9", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "BERT+theory-grounded SRL映射：CLS对role vectors的internal-state attention和ARG0→target directional external-cue attention，用learned gate融合context/role representation。", "evaluation_evidence": "四SEPF task officialtrain/dev/test，dev checkpoint选择与gridsearch；sentiment增益marginal非显著，emotion/politeness/formality较稳。", "caveats": [ "需实际SRL解析/role attention而不是附加role词袋", "mentalhealth数据应用不提供医疗有效性证据；sentiment负结果必须保留" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13797v1#Sx3", "https://arxiv.org/html/2610.13797v1#Sx4", "https://arxiv.org/html/2610.13797v1#Sx5" ], "evidence_sections": [ { "section_index": 2, "heading": "Theory-Grounded Role Mappings", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4997, 5397 ] ] }, { "section_index": 3, "heading": "REFLECT Framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9915, 10315 ] ] }, { "section_index": 4, "heading": "Experiments and Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5764, 6164 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13797.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13790", "assignment_index": 25, "title": "OccludeBench: Can VLMs Find What's Missing?", "published": "2026-10-09T16:04:54Z", "authors": [ "Jasin Cekinmez", "Addison J. Wu", "Ryo Mitsuhashi", "Linrong Cai", "Xingyu Fu", "Zhuang Liu" ], "affiliation": [ "Jasin Cekinmez † † thanks: Equal Contribution Addison J. Wu 1 1 footnotemark: 1 Ryo Mitsuhashi 1 1 footnotemark: 1 Affiliation: Linrong Cai Xingyu Fu Zhuang Liu Affiliation: Princeton University" ], "source_url": "https://arxiv.org/html/2610.13790v1", "source_sha256": "d2c24230c46efe4acc58b06d61d9c18e6b026e38551b847ef03c151c7c1c0671", "source_bytes": 245640, "source_json_sha256": "89f1ba1f9105fd5625917395ccfb5ce2baffbf5484926f5100135133b3338e03", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "人工完全occlude对象，direct physical-cue与inferentialcontext两类；OccludeTrace四阶段诊断；实际image-generation inpaint candidate再与原图cue confront的干预。", "evaluation_evidence": "五VLM与human free-form，direct/inferential差距；reason/tool/fewshot无稳定消除，只有两imagegen模型尝试inpaint。", "caveats": [ "不显著gap不是统计等价/证明已完全消除", "图片原始gold对象只能外部评分，不能给inpaint/confront prompt；需独立选validation策略" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/addisonwu05/OccludeBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/afs07fda89sdfas90/data", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13790v1#S2", "https://arxiv.org/html/2610.13790v1#S3" ], "evidence_sections": [ { "section_index": 1, "heading": "Section 2 OccludeBench", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4179, 4579 ] ] }, { "section_index": 2, "heading": "Section 3 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11431, 11831 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13790.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13783", "assignment_index": 26, "title": "UkrMeme: Multimodal Explainable Persuasion Detection in Ukrainian Memes", "published": "2026-10-09T16:03:27Z", "authors": [ "Nazar Andrushko", "Vera Schmitt", "Veronika Solopova" ], "affiliation": [ "Nazar Andrushko Affiliation: Ukrainian Catholic University, Lviv, Ukraine Vera Schmitt Affiliation: Johannes Gutenberg University Mainz, Mainz, Germany Affiliation: German Research Center for Artificial Intelligence (DFKI), Berlin, Germany Affiliation: Centre for European Research in Trusted AI (CERTAIN) nazar.checkemail@ucu.edu.ua vera.schmitt@uni-mainz.de veronika.solopova@uni-mainz.de Veronika Solopova Affiliation: Johannes Gutenberg University Mainz, Mainz, Germany Affiliation: German Research Center for Artificial Intelligence (DFKI), Berlin, Germany" ], "source_url": "https://arxiv.org/html/2610.13783v1", "source_sha256": "c778a577218948beaa248eaff04ade8ffff0f88343d90e97fc50e146e1bfd958", "source_bytes": 276918, "source_json_sha256": "4daa6f3c62978e61f43ccf418b2ee707b8466f83f500aadaea520941a3fe8c62", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "200乌克兰wartime memes新标注及Propaganda950翻译inpaint/OCR文本替换数据；zero-shot三modality hierarchicalF1，LRP对照modeltokenattribution。", "evaluation_evidence": "四模型和训练label频率weightedrandom baseline，explanationtoken category由GPT4o分类并小样本人核。", "caveats": [ "主要是领域数据/zero-shot经验分析，无本批foundation新训练或推理机制", "翻译/图文编辑及OCR噪声、政治冲突语境外推需谨慎，作为数据来源记录而非新adapter" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/JaidedAI/EasyOCR", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/eingrid/persuasion-techniques-ua-memes", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/enesmsahin/simple-lama-inpainting", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13783v1#S3", "https://arxiv.org/html/2610.13783v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11213, 11613 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15244, 15644 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13783.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13697", "assignment_index": 31, "title": "Improving Semantic Coverage in AI Test & Evaluation through Generative Augmentation", "published": "2026-10-09T14:57:12Z", "authors": [ "Connor Hashemi", "Bharadwaj Ravichandran", "Connor Anderson", "Shitij Govil", "Brandon RichardWebster", "Brian Hu", "Keith Fieldhouse" ], "affiliation": [ "Connor Hashemi Bharadwaj Ravichandran Connor Anderson Shitij Govil Brandon RichardWebster Brian Hu Keith Fieldhouse Affiliation: [" ], "source_url": "https://arxiv.org/html/2610.13697v1", "source_sha256": "4a717c5ca6768283ba69f3f06badae833b3990a7b1a14d3fff065d4f7b5c7b6d", "source_bytes": 292113, "source_json_sha256": "6cf7ec0bda2ad86a5f1465503ed1744d7e0b6b457ca78ff4178d2e89ba392506", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "zero-shot VLM二元semantic因子+embedding clustering辨识coverage gap，diffusion image2image迁移语义保content，再按embedding drift/radius/completeness验证合格batch加入test；不重训被评模型。", "evaluation_evidence": "CCT trailcam与BDD100K day/night toy gap，original/augmented/ideal真night测试对trained-day与diverse-trained模型区分。", "caveats": [ "合成增强是T&E诊断，不是独立自然分布test或能力训练", "真实VLM/diffusion和质量验证必需；改变test后比较需固定预注册coverage criterion" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Hashe037/DEDUCE", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13697v1#S3", "https://arxiv.org/html/2610.13697v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Overall Pipeline", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12020, 12420 ] ] }, { "section_index": 3, "heading": "4 Pipeline Results on Toy Datasets", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13697.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13690", "assignment_index": 32, "title": "Transformer-based In-Context Score Operator Learning for Diffusion Models", "published": "2026-10-09T14:53:49Z", "authors": [ "Wuzhe Xu", "Zhaiming Shen", "Yulong Lu", "Wenjing Liao", "Rongjie Lai" ], "affiliation": [ "Wuzhe Xu Affiliation: South Dakota State University Email: wuzhe.xu@sdstate.edu Zhaiming Shen Affiliation: Georgia Institute of Technology Email: zshen49@gatech.edu Yulong Lu Affiliation: University of Minnesota Email: yulonglu@umn.edu Wenjing Liao Affiliation: Georgia Institute of Technology Email: wliao@gatech.edu Rongjie Lai Affiliation: Purdue University Email: lairj@purdue.edu" ], "source_url": "https://arxiv.org/html/2610.13690v1", "source_sha256": "4f3625ff9247d3cdaff7b6bd07df28ee77c366dddbf2e9fc75a7e1f8338ee317", "source_bytes": 1137926, "source_json_sha256": "d6ff06c2d0c181dbb936ee49699a11924e3653017808a8095232a9b421b01d44", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "context empirical samples+noisy query/time输入query-dependent transformer score operator；Gaussian conditional-mean/Nadaraya-Watson exact-representable构造，跨task amortized denoising scorematch，unseen任务reverse diffusion不更新参数。", "evaluation_evidence": "Navier-Stokes ID/OOD、ERA5/QuickDraw/Food101更多ID；复杂task退化为论文明确negative，theorem对指定time-modulated mean class而非任意Transformer。", "caveats": [ "不能用经验kernel estimator当已训练neuraloperator能力结果", "泛化边界/复杂task退化保留，当前bounded numerical正文未穷尽表格/训练selection" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/googlecreativelab/quickdraw-dataset", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/woodssss/In-context-learning-for-Diffusion-Models", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13690v1#S3", "https://arxiv.org/html/2610.13690v1#S4", "https://arxiv.org/html/2610.13690v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10269, 10669 ] ] }, { "section_index": 3, "heading": "4 Theoretical results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "5 Numerical results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13690.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13660", "assignment_index": 33, "title": "Pretraining and Test-Time Training Budgets Matter for ARC: A Controlled Study of VARC and TRM", "published": "2026-10-09T14:33:30Z", "authors": [ "Adrien Deliege", "Marc Van Droogenbroeck" ], "affiliation": [ "Adrien Deliège † † thanks: Corresponding author. Marc Van Droogenbroeck Affiliation: Montefiore Institute of Electrical Engineering and Computer Science, University of Liège, Liège, Belgium E-mail adrien.deliege@uliege.be" ], "source_url": "https://arxiv.org/html/2610.13660v1", "source_sha256": "7c53ccd87967ed3507bb1b17e1889e507d07c31a134af8954850bad9f5b876b2", "source_bytes": 169939, "source_json_sha256": "9133dd7078889c00afd0181d14a8de3519cdf4d4c916953eafc66909877e7da5", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "在共同inductiveARC protocol下分别固定VARC/TRM架构，把task variant数与total sample presentations分开控制；TTT只用每testtask demonstrations、queryinput做feedforward并inverse-transform/vote。", "evaluation_evidence": "5 pretrain×4 TTT budgets两模型40 pass@2结果；VARC scratch竞争力、TRM更依赖pretrain，TTTvariant数并非两模型都单调受益。", "caveats": [ "不能将预算差异当架构提升；test query答案不能进TTT", "两架构更广domain外推留future；oracle值只能诊断不能promotion" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/AdrienDeliege/arc-training-budgets/tree/main", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13660v1#S3", "https://arxiv.org/html/2610.13660v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Budget Definitions and Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7338, 7738 ] ] }, { "section_index": 3, "heading": "4 Results and Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6193, 6593 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13660.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13636", "assignment_index": 34, "title": "Benchmarking Answer-Grounding Consistency in Remote Sensing Visual Question Answering", "published": "2026-10-09T14:15:50Z", "authors": [ "Ping Guo", "Zhiqi Huang", "Xinran Li", "Linkun Yu", "Jingyin Pei", "Kamata Seiichiro" ], "affiliation": [ "Ping Guo Zhiqi Huang Xinran Li Linkun Yu Jingyin Pei Kamata Seiichiro \\sthanks Corresponding author" ], "source_url": "https://arxiv.org/html/2610.13636v1", "source_sha256": "8a18c9ad61f8c143cf65e27b2ca96166f8c145093a5a271e63936781b3750638", "source_bytes": 124210, "source_json_sha256": "8844d039db957f974d1cdc931be97057614b84ffdd73277fff2a769dafd08961", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "AESP归一VLM自己的answer并验证predictedgrounding覆盖objectbox center；exist/count/compare规则支持answer，reported ES_all/ES_correct/jointsuccess而不只IOU。", "evaluation_evidence": "15027 G-RSVQA七releasedVLM，无新增训练；answer extractor三选一400响应人核，groundingprompt不含以前生成answer。", "caveats": [ "gt objectannotations为外部evidence scoring，不可给模型", "answer extractor选择子集与finaltest隔离要进一步核，domain-specific遥感结论不能替代通用grounding" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/pingggg516/AESP-Evaluation", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13636v1#S3", "https://arxiv.org/html/2610.13636v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 AESP Evaluation Protocol", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5801, 6201 ] ] }, { "section_index": 3, "heading": "4 Experiment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2341, 2741 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13636.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13635", "assignment_index": 35, "title": "PANDA: Prior-guided Attentional Dual-path Architecture", "published": "2026-10-09T14:15:14Z", "authors": [ "Rishabh Jain", "Federico Belotti", "Stefano Coniglio", "Pietro Lio", "Stefano Fiorini" ], "affiliation": [ "Rishabh Jain Affiliation: University of Cambridge, UK Email: rj412@cst.cam.ac.uk Federico Belotti Affiliation: University of Bergamo, Italy Email: federico.belotti@unibg.it Stefano Coniglio Affiliation: University of Bergamo, Italy Email: stefano.coniglio@unibg.it Pietro Liò Affiliation: University of Cambridge, UK Email: pl219@cst.cam.ac.uk Stefano Fiorini Affiliation: Independent Researcher Email: s.fiorini1994@gmail.com" ], "source_url": "https://arxiv.org/html/2610.13635v1", "source_sha256": "fb54054a6c23653fe4cb6a5ce3ffb42a25a72e1b00ff202954a00d04fb839169", "source_bytes": 1188456, "source_json_sha256": "23d5229d0c5211161d2fc067d481aef3001100ef46b5c2a2ea643e573677836e", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "两阶段node→hyperedge→node HNN里main path插值receiver-normalizedattention与structuralincidence prior，auxiliary每sender sigmoid noncompetitive变换；还分析directedcomplexspectraloperator。", "evaluation_evidence": "四HNN/八nodeclassification datasets，normalisedaux同参消融含少数小例外。", "caveats": [ "本论文是通用hypergraph nodeclassification插件，非LLM/VLM/预训练foundation路径", "保留graph方法证据，不为本foundation范围新增adapter" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Stefa1994/GeDi-HNN", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/jainris/panda", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13635v1#S3", "https://arxiv.org/html/2610.13635v1#S4", "https://arxiv.org/html/2610.13635v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 PANDA: Prior-guided AttentioNal Dual-path Architecture", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Generalised Directed Spectral Convolution as Two-Stage Message Passing", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6195, 6595 ] ] }, { "section_index": 4, "heading": "5 Experimental Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13635.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13578", "assignment_index": 37, "title": "RARA: Robust Policy Fine-Tuning through Retrieval-Anchored Representation Alignment for Robotic Manipulation", "published": "2026-10-09T13:31:20Z", "authors": [ "Xuhui Kang", "Sung-Wook Lee", "Yan Zhuang", "Tianle Zhong", "Yen-Ling Kuo" ], "affiliation": [ "Xuhui Kang Sung-Wook Lee Yan Zhuang Tianle Zhong Yen-Ling Kuo Affiliation: University of Virginia Affiliation: {xuhui, dcs3zc, yyx3hd, fad3ew, ylkuo}@virginia.edu" ], "source_url": "https://arxiv.org/html/2610.13578v1", "source_sha256": "03cf8c75f9004e0c4ea694cc7a54bc1e7203af929c6846a03f8b48259332203e", "source_bytes": 373378, "source_json_sha256": "ea3aef8ef275695d2291f21e7b41bafc51f9a0c7707b8f2880ed91cda5ef873c", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "RARA从pretrain demos按taskintent/motion retrieval，stage1 target-representationcontrastivealignment到frozenpretrainfeature，stage2 BC finetune同时anchor retrieved representations；anchorloss不需要actionlabels。", "evaluation_evidence": "DiffusionPolicy/GR00T、RoboTwin/RoboCasa/realrobot13tasks ID/OOD；额外对retrieved demos行为克隆反而降成功。", "caveats": [ "真实policy actionrollout与retrievedpretrain数据是核心，featureMSE toy不能claimrobustrobot能力", "realrobot闭环不是production onlineAB，layer/retieval超参validation需下一批核" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13578v1#S3", "https://arxiv.org/html/2610.13578v1#S4", "https://arxiv.org/html/2610.13578v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Problem Definition", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1393, 1793 ] ] }, { "section_index": 3, "heading": "4 RARA: Retrieval-Anchored Representation Alignment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9837, 10237 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13578.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13763", "assignment_index": 28, "title": "Neurons++: A Unified Visual-Stream Framework for Image and Video Reconstruction from fMRI with Neural-to-Visual Interpretability", "published": "2026-10-09T15:49:07Z", "authors": [ "Haonan Wang", "Yuchen Wang", "Jingyu Lu", "Qixiang Zhang", "Lehan Wang", "Xuanqi Huang", "Xiaomeng Li" ], "affiliation": [ "HKUST; Shenzhen Loop Area Institute (official PDF author block page1)" ], "source_url": "https://arxiv.org/pdf/2610.13763v1", "source_sha256": "1218f23bc0b65fc3c20d749075a71d5bb1d08f87a32e45bd77073542e3ef69ec", "source_bytes": 37538078, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "official PDF fallback; non-Google/Meta", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "official PDF pages1..7 first5500 extracted characters each; selected interpretability/limitations passages pages8..11; not complete25pages", "defining_method": "image-fMRI multisubject pretrain→subject-specific image/video adaptation；ridge+residualMLP/CLIP+diffusionprior alignment，四stage serial temporaltransformer持续接原neuralrepresentation，依次mask/concept/scene/blurryview监督；unCLIP keyframe后modalityappropriateSDXL或T2V diffusion。", "evaluation_evidence": "CC2017/HCP三subjects视频、NSD subjects1/2/5/7 images；HCP为作者group-aware temporal自建split，非官方test；ROI/randomgroupcontrol与delta-fMRI qualitative direction probes。", "caveats": [ "统一指共享设计/接口，不是跨image/video联合训练单decoder", "per-response zscore后delta只能qualitative方向；强ROIboost非校准causal gain", "公共neuraldata/pretrained diffusion真实输入不可toy随机ROI代理；code/weights声明未独立核仓库可用" ], "online_ab": "no production-traffic A/B established in reviewed PDF passages; not abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/xmed-lab/NeuronsPlusPlus", "status": "official PDF declares code/model weights; repository contents not independently verified" } ], "evidence_anchors": [ "https://arxiv.org/pdf/2610.13763v1#page=3", "https://arxiv.org/pdf/2610.13763v1#page=5", "https://arxiv.org/pdf/2610.13763v1#page=7", "https://arxiv.org/pdf/2610.13763v1#page=8" ], "evidence_sections": [ { "format": "PDF", "pages": [ 1, 2, 3, 4, 5, 6, 7 ], "first_extracted_characters_per_page": 5500 }, { "format": "PDF", "pages": [ 8, 9, 10, 11 ], "scope": "reported selected term-centered interpretability/limitations passages; not full pages" } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13763.pdf", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13557", "assignment_index": 38, "title": "Probabilistic Ship Trajectory Forecasting using Transformer Models on AIS Data", "published": "2026-10-09T13:13:07Z", "authors": [ "Frederik E. T. Schöller" ], "affiliation": [ "Frederik E. T. Schöller † † thanks: This work was carried out while the author was with the Automation and Control Group, Department of Electrical Engineering, Technical University of Denmark, Kgs. Lyngby, Denmark. The author is now an independent researcher. Affiliation: Technical University of Denmark Email: frederikschoeller@gmail.com" ], "source_url": "https://arxiv.org/html/2610.13557v1", "source_sha256": "8fd9d9e988ae5211936c8144637c69866fc4926fc7451134f100b9215df3f845", "source_bytes": 254665, "source_json_sha256": "2ec98e1f76ceb0a7d3f789d4d7d8d8fc0ca725b059d040bb7f16161b0739c3fb", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "任务专用Transformer/Performer预测AIS位置quantiles；区域KDE历史heatmap经CNN进入网络，pinballloss输出0.025/0.5/0.975，10%训练holdout选收敛。", "evaluation_evidence": "月度1636220 vesseltrajectories与ARPA样constant/linear/Gaussianbaseline，heatmap帮助anticipate地理turn。", "caveats": [ "从头训练航迹预测，不是本范围预训练foundation/LLM/VLM算法", "下一独立交通轨迹track应再核时间/船舶/区域split；当前不新增foundation adapter" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13557v1#S3", "https://arxiv.org/html/2610.13557v1#S4", "https://arxiv.org/html/2610.13557v1#S5", "https://arxiv.org/html/2610.13557v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Data", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5467, 5867 ] ] }, { "section_index": 3, "heading": "4 Transformer type networks", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9918, 10318 ] ] }, { "section_index": 4, "heading": "5 Training Specifications", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2891, 3291 ] ] }, { "section_index": 5, "heading": "6 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11309, 11709 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13557.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13548", "assignment_index": 39, "title": "Auditing VLM Benchmarks from Existing Evaluation Runs", "published": "2026-10-09T13:05:22Z", "authors": [ "Denis Degtiarev", "Anna Gromova", "Sergey Katkov", "Artem Mikhailin", "Safina Mukhamedshina", "Albert Nasybullin", "Karina Shaikhutdinova", "Artem Tuzov" ], "affiliation": [ "Denis Degtiarev Affiliation: Innopolis University Email: d.degtiarev@innopolis.university Anna Gromova Affiliation: Innopolis University Email: an.gromova@innopolis.university Sergey Katkov Affiliation: Innopolis University Email: s.katkov@innopolis.university Artem Mikhailin Affiliation: Innopolis University Email: a.mikhailin@innopolis.university Safina Mukhamedshina Affiliation: Innopolis University Email: s.mukhamedshina@innopolis.university Albert Nasybullin Affiliation: Innopolis University Email: a.nasybullin@innopolis.university Karina Shaikhutdinova Affiliation: Innopolis University Email: k.shaikhutdinova@innopolis.university Artem Tuzov Affiliation: Innopolis University Email: a.tuzov@innopolis.university" ], "source_url": "https://arxiv.org/html/2610.13548v1", "source_sha256": "54bedf686d50e27a471011fc3b36f81d9e74c81b922bcefdf4ff0875c3cc96fd", "source_bytes": 137373, "source_json_sha256": "758d309d7a979b656d3aaba8c859bc806d08757a21fbe18d870ac457b4e64d9f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "VLMEvalKit保存预测/判分按questionid对齐，新增no-image blindrun，六postprocessdetectors+matched-permutationnull/interval/modelpool敏感性；gold只外部audit。", "evaluation_evidence": "MMStar三模型和binary format calibration边界；曾goldbackfill把unanimousconflicts134压成12，已拒绝这种补值。", "caveats": [ "20.2%no-gain为pool aggregate，扣null excess才3.2pp；不证明每item无视觉或特定污染原因", "四选一基率不可直接迁移yes/no；需真实rawoutputs不可oracle补答案" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13548v1#S3", "https://arxiv.org/html/2610.13548v1#S4", "https://arxiv.org/html/2610.13548v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5747, 6147 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9208, 9608 ] ] }, { "section_index": 4, "heading": "5 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 647, 1047 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13548.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13527", "assignment_index": 40, "title": "S-Space: Exploring Spatial Workspace in Multimodal Language Models", "published": "2026-10-09T12:48:41Z", "authors": [ "Zhiyuan Shi", "Xiang Li", "Xingzhuo Guo", "Shangchen Miao", "Weiliang Chen", "Qiyu Dai", "Jinshan Ren", "Jiangran Lyu", "Fangfu Liu", "Mingsheng Long", "Jialong Wu" ], "affiliation": [ "MirroS", "Tsinghua University", "Peking University" ], "source_url": "https://arxiv.org/html/2610.13527v1", "source_sha256": "bbbbddef4d8d12075ae6f05e721de39297c1f69061ce6c576b67fb4264a7e77c", "source_bytes": 456358, "source_json_sha256": "d65be5287897a8d3ac1893d22e2c61128ed72ee2ad2d98a729c122e7c80a8b62", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "objecttoken activations的spatialanswer-score gradients识别水平/垂直/距离三轴linearS-Space，readout相对坐标margin，并用activation intervention检验verbal spatial因果使用。", "evaluation_evidence": "COCO轴识别、CV-Bench/EmbSpatial/SpatialTunnel不同prompts，MMSI viewpoint/reference-frame失败分解。", "caveats": [ "源于spatiallanguage训练是hypothesis不是受控training证明", "linearprobe/gradientreadout为诊断，不等于修复spatialcomputations；轴/layerselection须validation固定" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/mirros-lab/s-space", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3.6-27B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13527v1#S2", "https://arxiv.org/html/2610.13527v1#S3", "https://arxiv.org/html/2610.13527v1#S4", "https://arxiv.org/html/2610.13527v1#S5" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5446, 5846 ] ] }, { "section_index": 2, "heading": "3 S-Space as Internal Spatial Workspace", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4132, 4532 ] ] }, { "section_index": 3, "heading": "4 How Does S-Space Emerge? A Case Study", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2472, 2872 ] ] }, { "section_index": 4, "heading": "5 Diagnosing Spatial Reasoning Failures with S-Space", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6881, 7281 ] ] }, { "section_index": 6, "heading": "Authors", "read_character_ranges_in_cached_section": [ [ 0, 162 ] ] }, { "section_index": 7, "heading": "Affiliations", "read_character_ranges_in_cached_section": [ [ 0, 129 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13527.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13510", "assignment_index": 41, "title": "Retrieval-Augmented Multimodal Language Model with Heterogeneous Evidence Integration for Clinical Prognosis", "published": "2026-10-09T12:35:22Z", "authors": [ "Hamza Shafiq", "Jialu Tang", "Ting Dang", "Jun Hu", "Aaqib Saeed" ], "affiliation": [ "Hamza Shafiq m.h.shafiq@tue.nl Affiliation: Eindhoven University of Technology Jialu Tang j.tang@tue.nl Affiliation: Eindhoven University of Technology Ting Dang Ting.Dang@unimelb.edu.au Affiliation: University of Melbourne Jun Hu j.hu@tue.nl Affiliation: Eindhoven University of Technology Aaqib Saeed a.saeed@tue.nl Affiliation: Eindhoven University of Technology" ], "source_url": "https://arxiv.org/html/2610.13510v1", "source_sha256": "4f810eb2996c78c9e3d62b815fdb79b48f2076c1dbb0442fdd9822ab141749d2", "source_bytes": 607871, "source_json_sha256": "7d49b7186549c9c436645966eeec0b0cdf00433062bf242e5726a61cdfbb9b62", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "PRISM numericalfeaturetokens+ECGencoder+recencyweightedpriorECG与training-cohort dualmodalitypopulationmemory meanpooledretrieval embeddings，真实softtokens入LLM，不暴露individualretrievedrecords。", "evaluation_evidence": "MDS-ED1443tasks和外部MC-MEDsupervisedadapt；96hprearrival输入为baseline没有的证据，benchmark-window matched结果不胜最强baseline。", "caveats": [ "retrieval对random优势的paired sign p=.067非uniform证据；外部微调不是zero-shot外推", "fusion对照容量/参数/优化也变，未隔离LM效应；不提供临床安全部署结论" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/hamzashafiq28/PRISM", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13510v1#S3", "https://arxiv.org/html/2610.13510v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13510.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13411", "assignment_index": 42, "title": "In-Context Learning for System Identification: Probabilistic Formulation, Long-Sequence Handling, and Benchmark Validation", "published": "2026-10-09T11:21:14Z", "authors": [ "Matteo Rufolo", "Dario Piga", "Filippo Pura", "Marco Forgione" ], "affiliation": [ "Matteo Rufolo Affiliation: SUPSI, IDSIA - Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland Affiliation: USI, Faculty of Informatics, Lugano, Switzerland Email: matteo.rufolo@supsi.ch Dario Piga Affiliation: SUPSI, IDSIA - Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland Email: dario.piga@supsi.ch Filippo Pura Affiliation: SUPSI, IDSIA - Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland Email: filippo.pura@supsi.ch Marco Forgione Affiliation: SUPSI, IDSIA - Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland Email: marco.forgione@supsi.ch" ], "source_url": "https://arxiv.org/html/2610.13411v1", "source_sha256": "db3342fbf4f2bb2b17cdc4a8a7d674a2d57320ebe0d100e35bef857c2104c198", "source_bytes": 423497, "source_json_sha256": "75507c3607a740329a8adced75d87b2578cddcbc8ea3303de9ed03b0d47aea95", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "synthetic systemclass预训probabilisticTransformer encoder-decoder metamodel，recurrentpatch neuralnetwork处理长context；inference多窗inversevariance合成，uncertaintyaware合成数据可regularize传统parametricID。", "evaluation_evidence": "Wiener-Hammerstein syntheticpretrain与realnonlinearstandardbench/ OOD input-onlyadapt，RTX4090；部分传统模型RMSE略更优。", "caveats": [ "应真实systeminput/output与learnedprobabilisticoperator，不能toy线性回归假扮", "repo/weights声明指引用Pura2024资产，具体本paperrevision未核；不是所有datasets强于baselines" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13411v1#S3", "https://arxiv.org/html/2610.13411v1#S4", "https://arxiv.org/html/2610.13411v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Meta learning framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12285, 12685 ] ] }, { "section_index": 3, "heading": "4 Handling long sequences", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5365, 5765 ] ] }, { "section_index": 5, "heading": "6 Examples", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13411.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13370", "assignment_index": 43, "title": "Overcoming Data Imbalances with Expert Selection Biases", "published": "2026-10-09T10:45:19Z", "authors": [ "Ayush K. Tarun", "Clara Meister", "Antoine Bosselut" ], "affiliation": [ "Ayush K. Tarun\n Clara Meister\n Antoine Bosselut EPFL {ayush.tarun,clara.meister,antoine.bosselut}@epfl.ch † † thanks: Corresponding author." ], "source_url": "https://arxiv.org/html/2610.13370v1", "source_sha256": "878af635673c69d2e9fd60a9e766d752e255033ad3b64e5c59934a41a904376a", "source_bytes": 600333, "source_json_sha256": "212451f90d93c08e56c1beeac3939f20ef881eb123a69f4d9153fcbdaecc1c38", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "COAX按输入category给top-K expertselection加bias，使少数category偏向可过配容量的expert subset同时可跨组共享；改变池子容量不改变每tokenK=8 activecompute。", "evaluation_evidence": "多模态/多语/domain三MoE规模到10.6B，分配/alphasweep，CPT新语言Russian/Hindi略弱于vanilla但replay更好。", "caveats": [ "必须在真实MoE pretraining路由执行，离线expertcount模型不是能力实现", "基于category输入适用性/训练token等价只是对应实验scale，不能通用性能或native推理加速" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/MLCommons/peoples_speech", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13370v1#S3", "https://arxiv.org/html/2610.13370v1#S4", "https://arxiv.org/html/2610.13370v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4716, 5116 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1620, 2020 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13186, 13586 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13370.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13341", "assignment_index": 44, "title": "Understanding Mamba-3 Through Oscillatory Dynamics", "published": "2026-10-09T10:16:58Z", "authors": [ "Shutian Liang", "Yue Song" ], "affiliation": [ "Shutian Liang Yue Song † † thanks: Denotes corresponding author. Affiliation: College of AI, Tsinghua University" ], "source_url": "https://arxiv.org/html/2610.13341v1", "source_sha256": "0f61fb1aa043fc0f13193846c7be123aba91aefd9e2c90b41d8c4df5dfbbe4d2", "source_bytes": 584019, "source_json_sha256": "68c00f3b710f18185d188ff6a3d3da11b1d7e2c3755f17e52514a4f1bf84db09", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Mamba3 decay×inputdependentrotation的statepair等价underdampedsecondorderdynamics；控制相同state结构比critical/overdamped，保持读写phase一致并做pretrainedphaseintervention。", "evaluation_evidence": "state tracking、1.5BMamba3/SISO及语言pretrain长context/lengthgeneralization；learnedphase去除扰动退化。", "caveats": [ "homogeneous state转移解释不包含全部inputinjection，不能说证明完整Mamba普适性质", "需要真实SSM/LM和phase路径，harmonic toy仅mechanismdemo不能替代LM能力" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/datasets/cerebras/SlimPajama-627B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/state-spaces", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13341v1#S3", "https://arxiv.org/html/2610.13341v1#S4", "https://arxiv.org/html/2610.13341v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 An Oscillatory Interpretation of Mamba-3", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3132, 3532 ] ] }, { "section_index": 3, "heading": "4 Oscillation and State Tracking", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6414, 6814 ] ] }, { "section_index": 4, "heading": "5 Oscillation in Language Modeling", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11186, 11586 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13341.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13337", "assignment_index": 45, "title": "Exploring Attention Sinks in Text-to-Image Transformers from Diffusion and Autoregressive Models", "published": "2026-10-09T10:10:51Z", "authors": [ "Binglei Li", "Mengping Yang", "Zhiyu Tan", "Wenxu Qian", "Junping Zhang", "Hao Li" ], "affiliation": [ "Binglei Li Affiliation: Fudan University Affiliation: Shanghai Innovation Institute Mengping Yang Affiliation: Fudan University Zhiyu Tan Affiliation: Fudan University Affiliation: Shanghai Academy of AI for Science Wenxu Qian Affiliation: Fudan University Affiliation: Shanghai Innovation Institute Junping Zhang Hao Li Affiliation: Fudan University Affiliation: Shanghai Innovation Institute Project Lead   Corresponding Authors {blli24}@m.fudan.edu.cn, {jpzhang,lihao_lh}@fudan.edu.cn" ], "source_url": "https://arxiv.org/html/2610.13337v1", "source_sha256": "ee9c38c3ac95b3e899334b27bf7eb63f509cdf5a1359656478f5d6c57ff4248d", "source_bytes": 270174, "source_json_sha256": "3e19c6b1b7cda7b355f978fcce63a915c71dfc32dca4a795028377f90e89a7c6", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "按prompt/step attention cumulative mass定位T2I head/positionsinks，在FLUX/ZImage/Infinity上logit/value/head干预，并插registertokens再分配sinks。", "evaluation_evidence": "GenEval/DPG、HPSv2/ImageReward，H100；diffusion边界/corner vsAR首token/global，head稳定而positions随prompt变。", "caveats": [ "sink常重合modalityboundary/global位置，干预是pathwaysensitivity非独立sinkcausal效果", "register部分redistribution并非完全absorb，若仅headrank打分不算生成机制" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13337v1#S3", "https://arxiv.org/html/2610.13337v1#S4", "https://arxiv.org/html/2610.13337v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Analytical Framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7188, 7588 ] ] }, { "section_index": 3, "heading": "4 Systematic Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8897, 9297 ] ] }, { "section_index": 4, "heading": "5 Intervention Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8131, 8531 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13337.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13334", "assignment_index": 46, "title": "PDE-PFN: Prior-Data Fitted Neural PDE Solver", "published": "2026-10-09T10:08:51Z", "authors": [ "Jaehyeon Park", "Mingu Kang", "Dongseok Lee", "Woojin Cho", "Kookjin Lee", "Anthony Gruber", "Youngjoon Hong", "Noseong Park" ], "affiliation": [ "Jaehyeon Park 1 Mingu Kang 2 Dongseok Lee 3 Woojin Cho 4 Kookjin Lee 5 Anthony Gruber 6 Youngjoon Hong 3 1 1 1 Corresponding Authors Noseong Park 2 1 1 1 Corresponding Authors 1 SolverX, 2 KAIST, 3 Seoul National University, 4 TelePIX 5 Arizona State University, 6 Sandia National Laboratories jaehyeon.park@solverx.ai, sgn08050@kaist.ac.kr, lorafa@snu.ac.kr, woojin.py@gmail.com, kookjin.lee@asu.edu, adgrube@sandia.gov hongyj@snu.ac.kr, noseong@kaist.ac.kr" ], "source_url": "https://arxiv.org/html/2610.13334v1", "source_sha256": "ca725de41abe06938307377afc2a20339d43eb2fd52b4bd25606b5cc9d14cc73", "source_bytes": 900880, "source_json_sha256": "57313c6cc56d116dc1016131ecdda9d97ffb022b3d8a1fcbdcb517be04fc5b93", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "PDE-PFN以noisyapproxPINN solutions为prior，self/crossattention PFN对contextcoord/value与querycoord预测posteriorpredictivemean，训练含AGFreg并按contextsize缩放。", "evaluation_evidence": "2DCDR family无taskfinetune测试和SWE/CNSE crossPDE fine-tuning；DPOT原一步任务改四步影响baseline。", "caveats": [ "crossPDE结果需finetune不能宣称全zero-shot universalPDE solver", "真实PINNprior/learntPFN需要复现；修改baseline预测horizon导致与原paper数字不可直接比" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13334v1#S3", "https://arxiv.org/html/2610.13334v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8867, 9267 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13334.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13321", "assignment_index": 47, "title": "Host-Aware Web Crawling for LLM Pretraining Data", "published": "2026-10-09T10:03:04Z", "authors": [ "Malte Ostendorff" ], "affiliation": [ "Malte Ostendorff Affiliation: Common Crawl Foundation Email: malte@commoncrawl.org" ], "source_url": "https://arxiv.org/html/2610.13321v1", "source_sha256": "7ff8201e10614e1ef87a893aa92e6edd3e1d2b0edc3a702b59f8296bc72eddc0", "source_bytes": 316833, "source_json_sha256": "ea65d549294fabdf91059889b60c9781846581974b517c654e1ed00b7ddd0f89", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "从WARC质量classifier/hostgraph/coverage估计host间ICC，offline模拟先sample少数pages、按hostsamplequality决定crawl depth，而非只用harmoniccentrality分预算。", "evaluation_evidence": "CC2024-10约0.25%WARC classifiedsample与另fullcrawl预计算hostartifact两种regime；centrality某classifier不如random。", "caveats": [ "收益为offlinecrawl simulation，未执行productioncrawlerAB", "knownuncrawled估计假设同host观测density代表未爬pages；下一批只离线公共data评测，livecrawl另需authority" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/datasets/commoncrawl/gneissweb-annotation-host-testing-v1", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/commoncrawl/host-index-testing-v2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/kenhktsui/fineweb-edu-fasttext-classifier", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13321v1#S3", "https://arxiv.org/html/2610.13321v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Data and Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8231, 8631 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10677, 11077 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13321.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13276", "assignment_index": 48, "title": "CopRoute: Dependence-Aware Conformal Routing for Language Models", "published": "2026-10-09T09:27:47Z", "authors": [ "Siva Rajesh Kasa", "Santhosh Kumar Kasa", "Karan Gupta" ], "affiliation": [ "Siva Rajesh Kasa † † thanks: Equal contribution. † † thanks: This work does not relate to the author’s position at Amazon. † † thanks: Corresponding author. Email: kasa@u.nus.edu Santhosh Kumar Kasa 1 1 footnotemark: 1 2 2 footnotemark: 2 Karan Gupta" ], "source_url": "https://arxiv.org/html/2610.13276v1", "source_sha256": "c6595f73df717e2e6451d6cfbb03677bd208211323e0f77044200425fa1c19b5", "source_bytes": 900247, "source_json_sha256": "1836718922a28a1ccdb58733fa356540244f8ea0fccc338eeef2e2cd2dae7d89", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "designcorrectnessmatrix用greedytally按新增覆盖排序，frozen query stopping score形成nestedmodelsets；独立split calibration选最小certifiedcutoff，否则NON-CERTIFIED。", "evaluation_evidence": "RouterBench/RouterEval matchedall-missrisk；GSM8K单uniqueprompt、MBPPfullpoolmiss过高被排除；directtopmodel有更低cost但不能certify。", "caveats": [ "factor4为design fixedorder保证，cutoffoptimal只同一个frozenfamily；需exchangeability", "只保证selectedset至少一正确，不允许用testgold选输出或runtime first-correct oracle" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/kasakh/coproute", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13276v1#S3", "https://arxiv.org/html/2610.13276v1#S4", "https://arxiv.org/html/2610.13276v1#S5", "https://arxiv.org/html/2610.13276v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Dependence-aware routing", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6750, 7150 ] ] }, { "section_index": 3, "heading": "4 Theory", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6722, 7122 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13497, 13897 ] ] }, { "section_index": 5, "heading": "6 Limitations and Conclusion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3594, 3994 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13276.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13272", "assignment_index": 49, "title": "OWLET: On-Policy Weighted Learning via Evidence Transfer for Video Large Language Model Post-Training", "published": "2026-10-09T09:23:21Z", "authors": [ "Qitan Lv", "Ke Fan", "Hong Wang", "Wen Wu", "Yilin Wang", "Bowen Zhou", "Feng Wu", "Chao Zhang" ], "affiliation": [ "Qitan Lv † † thanks: This work is done when Qitan Lv is an intern in Shanghai AI Laboratory. Affiliation: University of Science and Technology of China Affiliation: Shanghai AI Laboratory Email: qitanlv@mail.ustc.edu.cn Ke Fan Affiliation: Shanghai AI Laboratory Affiliation: Tsinghua University Email: wanghong1700@mail.ustc.edu.cn Hong Wang Affiliation: University of Science and Technology of China Email: fengwu@ustc.edu.cn Wen Wu Affiliation: Shanghai AI Laboratory Email: fanke1@pjlab.org.cn Yilin Wang Affiliation: Shanghai AI Laboratory Email: wuwen@pjlab.org.cn Bowen Zhou Affiliation: University of Science and Technology of China Affiliation: Shanghai AI Laboratory Affiliation: Tsinghua University Email: wangyilin@pjlab.org.cn Feng Wu Affiliation: University of Science and Technology of China Email: zbw@pjlab.org.cn Chao Zhang † † thanks: The corresponding author. Affiliation: Shanghai AI Laboratory Affiliation: Tsinghua University Email: zhangchao@pjlab.org.cn" ], "source_url": "https://arxiv.org/html/2610.13272v1", "source_sha256": "5f6f11e5cd2c4e3a22b7c5cc994d1affb5086049d5f8089599b49c95965565f8", "source_bytes": 1063414, "source_json_sha256": "60230328593f9af31906b0e64d24828885116747e6919c724f711395c9330745", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "OWLET student用uniformframes自己on-policy rollout，same-size teacher对同response用global+questionlocalprivilegedview评分；tokenlogprobcontrast只重分positiveadvantage，nonpositive不变；verifieraccepted响应cosinehiddenstate alignment。", "evaluation_evidence": "Video-Evidence96K训练、四Qwen3.5scale五video benchmarks，预算/teacher/ECR/SSA ablations及其他Qwen3VL外推。", "caveats": [ "真实video teacherprivilege限训练/分析，evalstudent不能见answer-conditionedgoldkeyframes", "需实际policyloss/tokenweight/teacherstates，trajectoryreward-only替代不成立；bounded正文非全附录核验" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13272v1#S3", "https://arxiv.org/html/2610.13272v1#S4", "https://arxiv.org/html/2610.13272v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Motivated Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5284, 5684 ] ] }, { "section_index": 3, "heading": "4 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8925, 9325 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7540, 7940 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13272.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13263", "assignment_index": 50, "title": "SheafPrune: Sheaf-Guided Graph Propagation for Visual Token Pruning", "published": "2026-10-09T09:15:08Z", "authors": [ "Xiaoyang Ma", "Chi Wang", "Hong Sun" ], "affiliation": [ "Xiaoyang Ma 1 1 footnotemark: 1 Affiliation: Key Lab of Smart Agriculture Systems, Ministry of Education Affiliation: China Agricultural University Affiliation: Beijing, 100083, China Email: s20243081931@cau.edu.cn Chi Wang † † thanks: These authors contributed equally to this work. Affiliation: PHOENIX Lab Affiliation: Hangzhou Dianzi University Affiliation: Hangzhou, 310018, China Email: conanben77@hdu.edu.cn Hong Sun † † thanks: Corresponding author. This work was supported by Central Guidance Fund for Local Scientific and Technological Development Projects in Inner Mongolia (2024ZY0145) Affiliation: Key Lab of Smart Agriculture Systems, Ministry of Education Affiliation: China Agricultural University Affiliation: Beijing, 100083, China Email: sunhong@cau.edu.cn" ], "source_url": "https://arxiv.org/html/2610.13263v1", "source_sha256": "e171584ffbb0be90318415acaa2d35af54d5534e81046e8b57a8046a453dfded", "source_bytes": 1229875, "source_json_sha256": "9ad276309ab77dfb22f8036dedd1d50c707ad27567e186c1dae03bf9173d7081", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SheafPrune modalitymaps+sparse textvisualgraph传relevance，独立visualgraph propagate并node-degree纠正，ratio-conditioned4.5Mselector经frozenVLM语言loss训练，predecoder一次hardtopk，structuralnewline不占budget。", "evaluation_evidence": "四backbone十benchmark七ratio unifiedbaselines；selector之后prefill缩短，visualencoding/projection已发生；depth增加并非普遍更好。", "caveats": [ "必须训练可导selector与支持约束graph，简单attentiontopk非核心", "5.15x为论文指定netprefill系统测量，不能当A100本地receipt；threshold与quantile参数不混称" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13263v1#S3", "https://arxiv.org/html/2610.13263v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12518, 12918 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15126, 15526 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13263.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13262", "assignment_index": 51, "title": "Dual-Stream Memory for Vision-Language-Action Policies", "published": "2026-10-09T09:14:59Z", "authors": [ "Yang Wang", "Ping Liu", "Ruiyu Wang", "Xuhua Chen", "Lingfeng Zhang", "Tao Zhang" ], "affiliation": [ "Yang Wang 1 , Ping Liu 1 , Ruiyu Wang 1,2 , Xuhua Chen 1 , Lingfeng Zhang 1 , and Tao Zhang 1 2 Affiliation: 1 Magic-Lab Team, Magiclab Robotics Inc. 2 University of California, Los Angeles, USA" ], "source_url": "https://arxiv.org/html/2610.13262v1", "source_sha256": "e6267822f2ea9cc24ac0975342d2ff4c588109b90ca4f709f37e8998af441ff9", "source_bytes": 168430, "source_json_sha256": "bf236b964199c84b77ff61a6873ad33f36381cd2aff4d8a307bb116b0809f0b4", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "taskagnosticGatedDeltaNet把history成1temporal token，progress+actionconditionedfeature reconstruction监督；taskconditionretrieverfixedthreshold选真实frames，linear multiscalefusion16spatialtokens；冻结compressor后jointpolicyretrieval训。", "evaluation_evidence": "16RoboMME/4realrobot，50episodes/task×三evalseed；cup-swap Shell比baseline差，identitytracking未解决。", "caveats": [ "baselinepublishedcheckpointaggregation不同，整体成功率差非严格matchedcausal比较", "必须真实memory-dependentpolicyrollouts，历史摘要加caption不可替代；realrobot非productionAB" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13262v1#S3", "https://arxiv.org/html/2610.13262v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "III METHOD", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10633, 11033 ] ] }, { "section_index": 3, "heading": "IV EXPERIMENTS", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7239, 7639 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13262.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13261", "assignment_index": 52, "title": "Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers", "published": "2026-10-09T09:11:08Z", "authors": [ "Zhenke Duan", "Xin Li", "Jiqun Pan", "Dong Xiaofei", "Hanwen Ning", "Xinyuan Song" ], "affiliation": [ "Zhenke Duan Affiliation: School of Statistics and MathematicsZhongnan University of Economics and Law, Wuhan, China Affiliation: CodeSoul.Co, Hangzhou, China Xin Li Affiliation: School of Statistics and MathematicsZhongnan University of Economics and Law, Wuhan, China Jiqun Pan Affiliation: School of Statistics and MathematicsZhongnan University of Economics and Law, Wuhan, China Dong Xiaofei Affiliation: School of Statistics and MathematicsZhongnan University of Economics and Law, Wuhan, China Hanwen Ning † † thanks: Corresponding author. Affiliation: School of Statistics and MathematicsZhongnan University of Economics and Law, Wuhan, China Affiliation: Innovation and Talent Base for Digital Technology and Finance, ZUEL Xinyuan Song Affiliation: Department of StatisticsThe Chinese University of Hong Kong, Hong Kong SAR, China" ], "source_url": "https://arxiv.org/html/2610.13261v1", "source_sha256": "0758e23c79425613df5f7da45253209d5037fac5f570cef0743fac36cc4d71be", "source_bytes": 517740, "source_json_sha256": "86d12dd8ebcb2a87f47846459506a7cb8f9dc3dcd47d908e50c045e2e364240f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "DAMCHA input-dependentfullcross-headMahalanobis-like矩阵保留offdiagonalhead交互，stack-wise参数共享；数学严格扩展仅pre-softmaxscores在generator表达力/curvature假设下。", "evaluation_evidence": "小decoder CIFARpixelAR/WMT14/CommonGen与Enwik8/Text8/DCLM-Mix LM100Ksteps，heldoutvalidationPPL，不是frontier规模生产serving。", "caveats": [ "不宣称任意矩阵为PSD距离metric；下一批须深核metricgenerator及causalprefix/缓存一致性", "validationPPL是论文所报，不得伪称isolatedtest；理论riskbound不等于实际全模型性能保证" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/erwinmsmith/DAMCHA", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/afmck/text8", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/reformer-enwik8", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13261v1#S3", "https://arxiv.org/html/2610.13261v1#S4", "https://arxiv.org/html/2610.13261v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 DAMCHA", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8543, 8943 ] ] }, { "section_index": 3, "heading": "4 Theoretical View", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 635, 1035 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13261.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13259", "assignment_index": 53, "title": "Beyond a Single Crop: Contextualized Embeddings for Player Re-Identification", "published": "2026-10-09T09:07:45Z", "authors": [ "Ge Luo", "Changhuo Yang", "Weisi Fan", "Rui Yang", "Shaokai Wang" ], "affiliation": [ "Ge Luo Changhuo Yang Weisi Fan Rui Yang Shaokai Wang \\corresponding" ], "source_url": "https://arxiv.org/html/2610.13259v1", "source_sha256": "a937d379d3641f36fa389d7b72da20717c19fe6e4f929468409d9a11d5b66ddf", "source_bytes": 150692, "source_json_sha256": "1a805da76f54afc3a187d43bdaee057eff5cbc8d9c1049d0ee313a053758e38f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ContextReID用同frame co-occurring crops+precedingframes视觉context生成单playerembedding；contextuallearning/retrievalloss，推理只detections/frame membership无personID/tracklets。", "evaluation_evidence": "SportsMOT/DanceTrack/SoccerNet GSR三MOT转cropretrieval Rank1/mAP；t-SNE仅qualitativecohesion。", "caveats": [ "trackIDs可train/eval但不可input或构建contextoracle身份组", "specializedsportsretrieval不是全trackingpipeline或生产A/B，contextdetections协议需冻结" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13259v1#Sx3", "https://arxiv.org/html/2610.13259v1#Sx4" ], "evidence_sections": [ { "section_index": 2, "heading": "Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8723, 9123 ] ] }, { "section_index": 3, "heading": "Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11511, 11911 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13259.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13253", "assignment_index": 54, "title": "Tabular Foundation Models Are Effectively Shallow", "published": "2026-10-09T09:01:42Z", "authors": [ "Irene Cannistraci", "Julia E. Vogt" ], "affiliation": [ "Irene Cannistraci Julia E. Vogt Affiliation: Department of Computer Science Affiliation: ETH Zurich Affiliation: Switzerland Email: irene.cannistraci@inf.ethz.ch" ], "source_url": "https://arxiv.org/html/2610.13253v1", "source_sha256": "195239740656e8b04555e9a18615f9a16620418630700c6e1f177bc68ac017c0", "source_bytes": 935529, "source_json_sha256": "6cb65d302c43236635ffcd0e6e0eef744c01b33f6b85e58dda5083915185e3c8", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "用unlabeled calibration internalHs→He ridgeclosedform affine translator替换contiguousblockwindow，而非identityskip；teacheroutputKL在calibration选择最大压缩窗口。", "evaluation_evidence": "8TFM/171classificationdataset、临床/回归补充；RTX5090 GPU/CPU效率，oneblockregressors通常不保ranking。", "caveats": [ "label-free translator不意味着TFM标准context不含训练labels", "classification近无损不迁移所有regression；dual-axis affine overhead/encoder瓶颈影响实际加速" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/icannistraci/shallow-tfm", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13253v1#S3", "https://arxiv.org/html/2610.13253v1#S4", "https://arxiv.org/html/2610.13253v1#S5", "https://arxiv.org/html/2610.13253v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6164, 6564 ] ] }, { "section_index": 3, "heading": "4 Locating Computational Redundancy", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4388, 4788 ] ] }, { "section_index": 4, "heading": "5 Evaluation of Compressed Models", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6296, 6696 ] ] }, { "section_index": 5, "heading": "6 Efficiency", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3252, 3652 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13253.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13252", "assignment_index": 55, "title": "Vision Transformers Learn Word Meaning from Images Alone", "published": "2026-10-09T09:00:08Z", "authors": [ "Anders Gjølbye", "Sebastian Ray Mason", "Stella Frank", "Lars Kai Hansen" ], "affiliation": [ "Anders Gjølbye † † thanks: Equal contribution. Sebastian Ray Mason † † footnotemark: Stella Frank Lars Kai Hansen Computer Science Affiliation: Technical University of Denmark Email: agjma@dtu.dk Note: Code is available at gjoelbye/vision-word-semantics ." ], "source_url": "https://arxiv.org/html/2610.13252v1", "source_sha256": "ba2777699f47685a586ecc144eedf0655f596465a44f0eb7583963f3b60ad0ff", "source_bytes": 924026, "source_json_sha256": "7ff1ee70e1e981201662df0673999ac4480b0d22cd68140d8f5a35fc6d33701b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "真实visionencoder对renderedwordimages减matchedcontrol成worddirection，控制拼写/字体并比较nearestsemanticneighbors、countrycapitaloffset和visualclasscosine；另text/speechcontrastivefinetune。", "evaluation_evidence": "14models，多大DINOv3有结构而其他largeSSL不呈现；DINOH比7Bwordclass对齐强原因未解释。", "caveats": [ "image-only无caption监督不等于训练images无可读文字；不能夸称通用人类语义理解", "word-rendering诊断非日常识字能力benchmark，finetune多数其他visionmodels无同等收益" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/gjoelbye/vision-word-semantics", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13252v1#S3", "https://arxiv.org/html/2610.13252v1#S4", "https://arxiv.org/html/2610.13252v1#S5", "https://arxiv.org/html/2610.13252v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14928, 15328 ] ] }, { "section_index": 3, "heading": "4 Lexical Structure", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3861, 4261 ] ] }, { "section_index": 4, "heading": "5 Word-Class Alignment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 821, 1221 ] ] }, { "section_index": 5, "heading": "6 Alignment and Finetuning", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2234, 2634 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13252.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13239", "assignment_index": 56, "title": "Jag: Direct Box Prediction for Efficient Visual Grounding", "published": "2026-10-09T08:54:42Z", "authors": [ "Xiuyuan Zhu", "Ke Lu", "Hao Wu", "Zijin Du", "Dongming Zhang", "Jian Xue" ], "affiliation": [ "Xiuyuan Zhu Affiliation: University of Chinese Academy of Sciences, Beijing, China Affiliation: State Key Laboratory of Communication Content Cognition, Beijing, China Ke Lu Affiliation: University of Chinese Academy of Sciences, Beijing, China Affiliation: Peng Cheng Laboratory, Shenzhen, Guangdong, China * Correspondence: xuejian@ucas.ac.cn Hao Wu Affiliation: University of Chinese Academy of Sciences, Beijing, China Affiliation: State Key Laboratory of Communication Content Cognition, Beijing, China Zijin Du Affiliation: University of Chinese Academy of Sciences, Beijing, China Dongming Zhang Affiliation: State Key Laboratory of Communication Content Cognition, Beijing, China Jian Xue Affiliation: University of Chinese Academy of Sciences, Beijing, China" ], "source_url": "https://arxiv.org/html/2610.13239v1", "source_sha256": "1ea9bb4e2146a06c219f15cdb7d215993b064ee1cb753fe9992258befbbb283a", "source_bytes": 282741, "source_json_sha256": "739fdc81f077f0a03ea85f1172499467b6b49eb2f85e7f1282f60b690331fe8f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Jag真实Qwen3.5-0.8B image+refexpression末个有效inputtoken hidden读轻regressionhead同预测4continuouscoords；headwarmup后geometryloss联合adapt语言backbone/visualmerger/head，无coordtokendecode。", "evaluation_evidence": "RefCOCO/+ /g officialsplit combinedtraining、AccIoU.5；base完整适配对比和head-only消融，含wrong-instance/extent失败。", "caveats": [ "视觉encoder/图文prefill成本仍在，不是跳过VLM或纯imageclassscorer", "必须几何训练和jointrepresentationadaptation；base未taskfinetune比较不能只归因替换outputhead" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/modelscope/evalscope", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13239v1#S3", "https://arxiv.org/html/2610.13239v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7634, 8034 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9438, 9838 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13239.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13208", "assignment_index": 57, "title": "How Much Computation Does Every Forecast Need? Structured Block Skipping in Two Time-Series Foundation Models", "published": "2026-10-09T08:26:14Z", "authors": [ "Max Bourgeat" ], "affiliation": [ "Max  Bourgeat Affiliation: Polytechnique Montréal Affiliation: Montréal, QC, Canada Email: max.bourgeat@polymtl.ca" ], "source_url": "https://arxiv.org/html/2610.13208v1", "source_sha256": "cb10c5248ff9601d6401cf949704418cb329ffc345298eb4df441e8c941de2e9", "source_bytes": 124004, "source_json_sha256": "d2bde114538a1e0b60e421b5108d74cb8c906b8b5e144bea6d87b7fcdb59c915", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "真实ChronosBolt/TImesFM2.5原生block早返回bypass，allblocks-enable exactbaseline；staticmasks只calibration选，executioncounters与syncGPU测成本。", "evaluation_evidence": "四series1088heldoutwindows A100，moderateaggregate近中性掩盖Traffic/Weather异质；equalbudgetperwindoworacle retrospectivelybetter但LODO routing失败。", "caveats": [ "oracle访问heldoutloss只能diagnostic，未形成可部署dynamic方法", "weights仍resident，不能宣称memory/energy节省；architecture/domain staticmask需calibration且不可用test选" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13208v1#S3", "https://arxiv.org/html/2610.13208v1#S4", "https://arxiv.org/html/2610.13208v1#S5", "https://arxiv.org/html/2610.13208v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Experimental design", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3144, 3544 ] ] }, { "section_index": 3, "heading": "4 Static redundancy yields real acceleration", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3022, 3422 ] ] }, { "section_index": 4, "heading": "5 Mask preference varies across windows", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2867, 3267 ] ] }, { "section_index": 5, "heading": "6 Discussion and limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 575, 975 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13208.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13154", "assignment_index": 58, "title": "From Glance to Focus: Mitigating Recognition Laziness in MLLMs for Fine-Grained Recognition", "published": "2026-10-09T07:42:32Z", "authors": [ "Yutong Yang", "Zhuqi Liu", "Mouxing Yang", "Xi Peng" ], "affiliation": [ "Yutong Yang Zhuqi Liu Mouxing Yang Xi Peng † † thanks: Y. Yang, Z. Liu, and M. Yang are with the College of Computer Science, Sichuan University, Chengdu 610065, China\n(e-mail: {yutongyang.gm, yunfree36, yangmouxing}@gmail.com).\nX. Peng is with the School of Artificial Intelligence, Sichuan University, Chengdu 610065, China,\nand also with Tianfu Jincheng Laboratory, Chengdu 610093, China\n(e-mail: pengx.gm@gmail.com)." ], "source_url": "https://arxiv.org/html/2610.13154v1", "source_sha256": "57c147fd6e17670b2c36f1b6e3fbfc20841b2417ad46c9af82ee9c997ba896f8", "source_bytes": 567265, "source_json_sha256": "a4503674c16b9d7e21c7db3c82f818c6b54fb1fa95dc70dc51fbe5da27d636d8", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "LIVE trainingfree以query视觉relevance+intercategoryaffinitygraph做adaptivecandidate inference，VLMselfreflection构建categoryspecificheaders/细粒度comparativeskills，再按uncertainty/confusion选择detailguidance。", "evaluation_evidence": "13 recognitionbench多backbone；fixedtopK部分Flower设置更差，compactcandidate必须兼顾coverage。", "caveats": [ "真实image候选与预建skill流程，不是把goldcategory放提示中", "candidate coverage是外部评估，code uponacceptance当前未确认release，validation选择待深核" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13154v1#S3", "https://arxiv.org/html/2610.13154v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "III Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "IV Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13154.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13128", "assignment_index": 60, "title": "Locate, Judge, Read: Role-Specialized Tool Agents over a Knowledge-Point Index for Open-Domain Multimodal Document QA", "published": "2026-10-09T07:21:36Z", "authors": [ "Wenlin Lin", "Fanzhe Wei" ], "affiliation": [ "Wenlin Lin Affiliation: Metask Lab Email: linwenlin.virhum@gmail.com Fanzhe Wei Affiliation: Metask Lab Email: whyer1@gmail.com" ], "source_url": "https://arxiv.org/html/2610.13128v1", "source_sha256": "6e29c15ecace7d0caf92a86f5dff11c1829b724ec2c5b9796de0538fd4b9a04c", "source_bytes": 342910, "source_json_sha256": "0ab682209505245a8d3fb41f8e97a851e239f4e45226a36fe1cc27001392b5ec", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "KnowForge indexatomicpoints/hybrid retrieval，Locator只search，codebridgejudge强制二hop、typedboundedpool合并rank，DeepReader只读真实text/image/table，synthesizer+formatcheck。", "evaluation_evidence": "M3DocVQA/MMLongBenchDoc/MultiHopRAG fullruns和mid300sweeps；多数数字singlefullrun，视觉reader在textcorpus贡献0；selfhostserver共享productiontraffic仅说明latency环境。", "caveats": [ "shared productionload不是用户onlineAB效果；不同backbonepublishedbaseline不完全matched", "mid300与fullbenchmark策略选择隔离尚未完整核；hostedKB注册不能自动授权，下一批优先local公开index" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/metask-ai/knowforge-bench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13128v1#S3", "https://arxiv.org/html/2610.13128v1#S4", "https://arxiv.org/html/2610.13128v1#S5", "https://arxiv.org/html/2610.13128v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6552, 6952 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1786, 2186 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 5, "heading": "6 Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2449, 2849 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13128.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13111", "assignment_index": 61, "title": "What Do Visual-Head Discovery Scores Actually Find? Attention Source, Necessity, and Random-Control Pitfalls in Vision-Language Models", "published": "2026-10-09T07:06:17Z", "authors": [ "Md Abdul Kadir", "Daniel Sonntag" ], "affiliation": [ "Md Abdul Kadir & Daniel Sonntag Affiliation: German Research Center for Artificial Intelligence (DFKI) Affiliation: Saarbrücken, Germany Affiliation: Oldenburg University Affiliation: Oldenburg, Germany Affiliation: {abdul.kadir, daniel.sonntag}@dfki.de" ], "source_url": "https://arxiv.org/html/2610.13111v1", "source_sha256": "2c19e4920ab61aedcde569e7b7bd5eaf8fefd977a977a331ef6e0c3f29ee3bf8", "source_bytes": 757210, "source_json_sha256": "930b4199c2d88b5db9d4cd8b74e429e8c7a16cdea05aad9eea626bd26d3bd0c4", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "对knowntargetregion attentionmass排名，querytype/corpus/attention-source 2×3×2 discovery；steeringCERS与masknecessityNCS分别对layer及attentionmassmatchedrandomheadsets验证。", "evaluation_evidence": "Qwen3VL8B 12sets、250random核心对照与其他VLMreplication；steerability无特异优势，necessity按attention source明显不同。", "caveats": [ "known-region干预是mechanisticdiagnostic，不能把gtregion告知实际groundingagent", "single randomdraw无法支持headimportance；随机基线宽、rankp分辨率/多重校正必须保留" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13111v1#S3", "https://arxiv.org/html/2610.13111v1#S4", "https://arxiv.org/html/2610.13111v1#S5", "https://arxiv.org/html/2610.13111v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Discovery Scores and Causal Tests", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5986, 6386 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 802, 1202 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15441, 15841 ] ] }, { "section_index": 5, "heading": "6 Discussion and Recommendations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2718, 3118 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13111.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13106", "assignment_index": 62, "title": "Scaling RNA Foundation Models through Iterative Refinement", "published": "2026-10-09T06:59:15Z", "authors": [ "Pengan Chen", "Liang Hong", "Xinyuan Liu", "Han Huang", "Ziqian Lin", "Jiayang Chen", "Yimin Fan", "Jiyue Jiang", "Yu Li" ], "affiliation": [ "Pengan Chen * Liang Hong * Xinyuan Liu * Han Huang Ziqian Lin Affiliation: Jiayang Chen Yimin Fan Jiyue Jiang Yu Li † Affiliation: Department of Computer Science and Engineering, The Chinese University of Hong Kong Affiliation: {chenpengan,liang.hong,xinyuan}@link.cuhk.edu.hk , liyu@cse.cuhk.edu.hk" ], "source_url": "https://arxiv.org/html/2610.13106v1", "source_sha256": "2674df0b300e6f24153bd8d9c39677a4a8b86a35dd7a823e8e1c534c63d6fa20", "source_bytes": 391042, "source_json_sha256": "32d2f7216c26ef9ee8f45ba85db6d93cc4111b239fe9e7a876a080a59763e1cd", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "EcoRNA prelude→sharedinputconditionTransformercore迭代→coda，训练跨depth maskedreconstruction/state reg/one-sideddeep-loss以shallowreference锚定observednucleotide；参数不随K增长。", "evaluation_evidence": "80M RNAGym70assays、sequencefamily/structure/chemicalprobing；depthsweep固定checkpoint/head，coarselabel较早plateau而denseoutputs后续小增益。", "caveats": [ "fitness报告absoluteSpearman不等于方向准确；4iterations500x参数对照不同pretrainingbreadth", "须在独立validation选inferencedepth，不能testassay选K；supervisedtasks适配后再固定depthsweep" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13106v1#S2", "https://arxiv.org/html/2610.13106v1#S3", "https://arxiv.org/html/2610.13106v1#S4" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14434, 14834 ] ] }, { "section_index": 2, "heading": "3 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10692, 11092 ] ] }, { "section_index": 3, "heading": "4 Discussion and conclusion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1517, 1917 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13106.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13143", "assignment_index": 59, "title": "Measurable Semantic 3D Gaussian Splatting with Joint Region and Depth Supervision for Interactive Building Inspection", "published": "2026-10-09T07:33:37Z", "authors": [ "Yi Zhang", "Tian Tan", "Ray Kai Leung Su" ], "affiliation": [ "Department of Civil Engineering, University of Hong Kong (official PDF p1)" ], "source_url": "https://arxiv.org/pdf/2610.13143v1", "source_sha256": "9c45adc270721db3badb54c823e18dc2f0a6f11d2a6d6167f454c6a253b118c0", "source_bytes": 2795585, "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "official PDF fallback; domain scope manually checked", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "PDF pages1..6 first2300chars each; pages14,15,16,25,27,43,46 first1100chars each; not complete59pages", "defining_method": "MVS reference直接监督Gaussian center pointdepth而非opacity混合renderdepth，foregroundmask限制RGB/depthdomain；semanticfeature绑Gaussianprimitive，AEC SplatTwin确定性measurement平台由MLLM解释结果。", "evaluation_evidence": "QuakeCity synthetic和HKUPB smartphone；depthMAE对同MVS训练reference是几何一致性；独立window scale校准后7targets外部manualmeasurement。", "caveats": [ "核心是建筑3DGS几何/语义表示，MLLM仅界面/解释已有测量，不是本foundation模型机制候选", "depth与train同reference不能claimabsolutephysicsaccuracy；合成QuakeCity没有realworld尺寸，physicalmeasurement另验证" ], "online_ab": "no quantified production user-traffic A/B established in selected PDF passages", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/pdf/2610.13143v1#page=14", "https://arxiv.org/pdf/2610.13143v1#page=25", "https://arxiv.org/pdf/2610.13143v1#page=43" ], "evidence_sections": [ { "format": "PDF", "pages": [ 1, 2, 3, 4, 5, 6 ], "first_extracted_characters_per_page": 2300 }, { "format": "PDF", "pages": [ 14, 15, 16, 25, 27, 43, 46 ], "first_extracted_characters_per_page": 1100 } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13143.pdf", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13102", "assignment_index": 63, "title": "From 2.1M Hours of Raw Broadcast Audio to Finnish Speech SSL: Confidence Filtering and Architectural Analysis at Scale", "published": "2026-10-09T06:57:05Z", "authors": [ "Yaroslav Getman", "Tamás Grósz", "Tommi Lehtonen", "Mikko Kurimo" ], "affiliation": [ "Yaroslav Getman Affiliation: Dept. lnformation and Communications Engineering Aalto University Espoo, Finland yaroslav.getman@aalto.fi Tamás Grósz Affiliation: Walton Institute, South East Technological University, Ireland Aalto University Espoo, Finland tamas.grosz@setu.ie Tommi Lehtonen Affiliation: Finnish\nNational Audiovisual Institute (KAVI) Helsinki, Finland tommi.lehtonen@kavi.fi Mikko Kurimo Affiliation: Dept. lnformation and Communications Engineering Aalto University Espoo, Finland mikko.kurimo@aalto.fi" ], "source_url": "https://arxiv.org/html/2610.13102v1", "source_sha256": "70b1fe29163bf100c760e31e736b65de0d58120f0b092b99f7923258b36243fc", "source_bytes": 262451, "source_json_sha256": "0dc32b22a1ad7a032668dae00bd054a37372535ff5f003417879a4a7b3bfce4d", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Finnishwav2vec2 SSL用ASRconfidence筛pretrain音频而非仅VAD/LID存在判断，backendsTransformer/Macaron/Conformer/EBranchformer按参数与pretrain数据matched对照。", "evaluation_evidence": "2.1M原broadcasthours六ASRtestsets、95/317/965M模型，512MI250Xpretrain；部分Xlarge domain比Large退化，conv优势随容量减弱。", "caveats": [ "confidence是质量代理可能改变语言/方言覆盖，不能把筛选本身当保证无bias", "broadcast数据license/发布状态未独立核；原大规模训练远超本批resources，只下一批拟公共小诊断" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/aalto-speech/large-scale-monolingual-speech-foundation-models", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/blog/Pclanglais/youtube-commons", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13102v1#S3", "https://arxiv.org/html/2610.13102v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "III Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4148, 4548 ] ] }, { "section_index": 3, "heading": "IV Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9942, 10342 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13102.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13064", "assignment_index": 64, "title": "CB-Orchestrator: Adaptive Workflow Optimization for LLM Agents via Contextual Bandits", "published": "2026-10-09T06:25:19Z", "authors": [ "Jiahang Sun", "Zhiwei Shang", "Zhipiao Liu", "Hongwei Yang", "Guoqing Xie", "Shuang Qiu", "Zhongxiang Dai" ], "affiliation": [ "Jiahang Sun Affiliation: The Chinese University of Hong Kong, Shenzhen Zhiwei Shang Affiliation: The Chinese University of Hong Kong, Shenzhen Zhipiao Liu Affiliation: Huawei Technologies Hongwei Yang Affiliation: Huawei Technologies Guoqing Xie Affiliation: Huawei Technologies Shuang Qiu Affiliation: City University of Hong Kong Zhongxiang Dai † † thanks: Corresponding author. Correspondence to daizhongxiang@cuhk.edu.cn. Affiliation: The Chinese University of Hong Kong, Shenzhen" ], "source_url": "https://arxiv.org/html/2610.13064v1", "source_sha256": "47627ed17014e8d3a9883c79c595b555a8d01db42e07f84ea723fa5c8cd7ad8d", "source_bytes": 487327, "source_json_sha256": "bc45042cd6183d8dabeccbadcdadd3d4ce8a431edb84ff90888ff77ca15957b8", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "候选workflow为可执行asyncPythonclass/operator组合，evolutionarysearch建pool，再contextualbandit探索执行workflow收reward训练selector，infer无explorationbonus并Wilsonfilter未充分trialworkflow。", "evaluation_evidence": "GSM8K/MATH/HumanEval/MBPP/GAIA，training-token开销同时报告是否包括poolconstruction。", "caveats": [ "controller须真正execute/evaluate operator，metadata workflowlabel或gold lookup不是集成", "独立validation/test与生成代码sandbox必需；optionalonlineadapt不等于已证明productiononlineAB" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13064v1#S3", "https://arxiv.org/html/2610.13064v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10868, 11268 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11074, 11474 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13064.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13060", "assignment_index": 65, "title": "KV-Repurpose: A Testbed for Evaluating Non-Prefix KV-Cache Reuse", "published": "2026-10-09T06:17:10Z", "authors": [ "Taiga Ikeda", "Daisuke Miyashita", "Jun Deguchi" ], "affiliation": [ "Taiga Ikeda Daisuke Miyashita Jun Deguchi Affiliation: Kioxia Corporation, Japan" ], "source_url": "https://arxiv.org/html/2610.13060v1", "source_sha256": "b314b98efc2856dcd457c925492d574ccc6b86634649118f6422555ab3c7abdf", "source_bytes": 153624, "source_json_sha256": "5140414ea78fae70beab85df65bf3ebc5d35089af515adb08bcd800c065bf192", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "vLLM扩展server+cacheproxy做非prefixsegmentKV复用，positional correction、segmentschedule/promptcapacity/LRU/NVMe监控；pseudoRAGquestioncontextdatabase形成workload。", "evaluation_evidence": "GPTOSS120B单H200，LongBenchPro/LongBench与Zipf workload20%sharedprefix；crossdocument reasoning明显局限，cachehit/TTFT为servingloop。", "caveats": [ "简单positioncorrection非exact语义cache等价；不是所有models/workloads适用", "pseudoRAG非真实检索能力，人工Zipf非productiontrafficAB；必须真实KVcache path不能命中率模拟器代理" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/LMCache/LMCache/issues/3238", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/taiga333/eval-kv-repurpose", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/vllm-project/guidellm", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13060v1#S2", "https://arxiv.org/html/2610.13060v1#S4", "https://arxiv.org/html/2610.13060v1#Sx1" ], "evidence_sections": [ { "section_index": 1, "heading": "2 System Implementation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6164, 6564 ] ] }, { "section_index": 3, "heading": "4 Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6945, 7345 ] ] }, { "section_index": 5, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 681 ], [ 281, 681 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13060.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13056", "assignment_index": 66, "title": "A4-MR: Abnormality-aware Anatomical Alignment for MR", "published": "2026-10-09T06:10:10Z", "authors": [ "Noel D'souza", "Zhijian Yang", "Nishank Singla", "Pranay Doshi", "Krisztian Koos", "Fei Tan", "Bruno Nunes", "Laszlo Rusko", "Bernadett Kolozsvari", "Eszter Sabo", "Emanuele Valeriano", "Vikash Gupta", "Seyed Iman Zare Estakhraji", "Marc Lebel", "Farzin Haddadpour", "Taha Kass-Hout", "Erhan Bas" ], "affiliation": [ "Noel Dsouza Affiliation: GE Healthcare, Seattle, USA Affiliation: Corresponding author:\nNoel Dsouza; Erhan Bas, {noel.dsouza, erhan.bas}@gehealthcare.com Zhijian Yang † Affiliation: GE Healthcare, Seattle, USA Nishank Singla Affiliation: GE Healthcare, Seattle, USA Pranay Doshi † Affiliation: GE Healthcare, Seattle, USA Krisztian Koos Affiliation: GE Healthcare, Budapest, Hungary Fei Tan Affiliation: GE Healthcare, Budapest, Hungary Bruno Nunes Affiliation: GE Healthcare, Budapest, Hungary Laszlo Rusko Affiliation: GE Healthcare, Budapest, Hungary Bernadett Kolozsvari Affiliation: GE Healthcare, Budapest, Hungary Eszter Sabo Affiliation: GE Healthcare, Budapest, Hungary Emanuele Valeriano Affiliation: GE Healthcare, Seattle, USA Vikash Gupta Affiliation: GE Healthcare, Seattle, USA Affiliation: Amazon, USA Seyed Iman Zare Estakhraji Affiliation: GE Healthcare, Seattle, USA Marc Lebel Affiliation: GE Healthcare, Seattle, USA Farzin Haddadpour † Affiliation: GE Healthcare, Seattle, USA Taha Kass-Hout Affiliation: GE Healthcare, Seattle, USA Erhan Bas Affiliation: GE Healthcare, Seattle, USA Affiliation: Corresponding author:\nNoel Dsouza; Erhan Bas, {noel.dsouza, erhan.bas}@gehealthcare.com" ], "source_url": "https://arxiv.org/html/2610.13056v1", "source_sha256": "54f90c196665df175135b06eefb00a4d14a3e07c321446cd75902939e7d476bb", "source_bytes": 797135, "source_json_sha256": "a12859bab844c6c39fe858dd177489aad62bb80935dcc19d04d98607ef91645b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "A4MR 3Dvisionlanguagepretrain anatomicalnoisylocalization/image-textregion对齐，healthy同organ可share positives，abnormal同organ不同finding区分；不是普通wholevolumeCLIP。", "evaluation_evidence": "约50kMRIstudies/491kseries，下游frozenencoder localization/segmentation/pathologytasks；MedSAM2带gtboxes比较属于额外oraclecontext。", "caveats": [ "TotalSegmentator/专有segmodel/DICOM transfer/reportLLM弱标噪声仅子集人核", "publiccheckpoint与clinicaldataaccess未独立核；不提供临床部署可靠性，gtboxreference不能给被评无prompt模型" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13056v1#S3", "https://arxiv.org/html/2610.13056v1#S4", "https://arxiv.org/html/2610.13056v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "5 Limitations and Future Work", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2942, 3342 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13056.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13045", "assignment_index": 67, "title": "Redirecting Teacher Supervision: Temporal-Response-Guided Latent Distillation for Video-Language Models", "published": "2026-10-09T06:01:48Z", "authors": [ "Xinkai Lu", "Junli Zha", "Haolan Wei", "Jiahui Wang", "Chao Zhou", "Xiang Song" ], "affiliation": [ "Xinkai Lu Junli Zha Haolan Wei Jiahui Wang Chao Zhou Xiang Song Affiliation: SF Technology" ], "source_url": "https://arxiv.org/html/2610.13045v1", "source_sha256": "ad3f12364320a3e8d42d807abd6c73c427767a5a79477f49f9b1db5aee810697", "source_bytes": 239905, "source_json_sha256": "a0e0ca2cfa995848ac36827ec7f4f70da78fb6a86dc0fcb870e896c7872d61f1", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "TRG在acceptedteacherresponse固定context下用student局部temporalperturb/JVP response构orthoprojectorP；teacher correction δ→δ+Pδ再restore原norm，保留span外分量，detachedtarget只redirectdirection。", "evaluation_evidence": "VideoR1SFT+OPD、4B/9Bstudent27Bteacher四bench；与PlainAS/randomorientation对比，VideoMMMU多数来自plainalignment。", "caveats": [ "数学保持是构target时localcosloss/representationgradnorm，非全参数grad或convergence保证", "JVP/subspace增加traincompute，equalupdates未equalcompute；projection-only或单独normalizedPδ会改定义" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13045v1#S3", "https://arxiv.org/html/2610.13045v1#S4", "https://arxiv.org/html/2610.13045v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Temporal-Response Guidance", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11693, 12093 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10239, 10639 ] ] }, { "section_index": 4, "heading": "5 Discussion and Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1092, 1492 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13045.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13040", "assignment_index": 69, "title": "Lightweight Time-Frequency Vibration Sensing for Consumer Device Condition Monitoring", "published": "2026-10-09T05:57:15Z", "authors": [ "Po-Heng Chou", "Jen-Yu Chiu", "Chun-Yu Yeh", "Wei-Lung Mao", "Walid Saad" ], "affiliation": [ "Po-Heng Chou 1,3 ,\nJen-Yu Chiu 2 ,\nChun-Yu Yeh 2 , Wei-Lung Mao 2 , and Walid Saad 3 Affiliation: 1 Research Center for Information Technology Innovation (CITI), Academia Sinica (AS), Taipei 11529, Taiwan 2 Department of Electrical Engineering, National Yunlin University of Science and Technology (NYUST), Yunlin 64002, Taiwan 3 Bradley Department of Electrical and Computer Engineering (ECE), Virginia Tech (VT), Alexandria, VA 22305, USA E-mails: d00942015@ntu.edu.tw, asd1248106@gmail.com, aa745579aa@gmail.com, wlmao@yuntech.edu.tw, walids@vt.edu" ], "source_url": "https://arxiv.org/html/2610.13040v1", "source_sha256": "10e04f2196460eee849f13963691ceffaa67e9283c8dd5d0f5771e3e31e9547b", "source_bytes": 131162, "source_json_sha256": "206d4b4abdfa7f39c8e80bc8bf3ba10b39903a4899bd1d4e9c8cfd1d77c642f2", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "CWT/HHT将振动变timefrequency图，再MobileNet/ConvNext/YOLO11 classifier；区别randomwindowsplit与bearing/timeblockstructuredtrain/val/test，记录preprocess+CPU网络成本。", "evaluation_evidence": "CWRU/PU/IMS bearingbench和i7-13700K batch1，非真实consumerappliance/embeddedMCU部署。", "caveats": [ "主要是任务专用信号分类而非本foundation模型方法", "structured隔离避免重叠window泄漏；不能将hostCPU加速推为洗衣机/风扇硬件表现" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.13040v1#S3", "https://arxiv.org/html/2610.13040v1#S4", "https://arxiv.org/html/2610.13040v1#S5", "https://arxiv.org/html/2610.13040v1#S7" ], "evidence_sections": [ { "section_index": 2, "heading": "III Proposed Framework and Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4511, 4911 ] ] }, { "section_index": 3, "heading": "IV Experimental Setup and Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4977, 5377 ] ] }, { "section_index": 4, "heading": "V Experimental Results and Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6710, 7110 ] ] }, { "section_index": 6, "heading": "VII Limitations and Future Work", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 314, 714 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13040.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.13021", "assignment_index": 70, "title": "EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents", "published": "2026-10-09T05:23:00Z", "authors": [ "Jie Huang", "Yanan Chen", "Ruixun Liu", "Kaichen He", "Xinyi He", "Rui Huang", "Zilong Zheng", "Zhenliang Zhang", "Yiwu Zhong" ], "affiliation": [ "Jie Huang Affiliation: School of Intelligence Science and Technology, Peking University Affiliation: State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China Yanan Chen Affiliation: School of Intelligence Science and Technology, Peking University Ruixun Liu Affiliation: School of Intelligence Science and Technology, Peking University Kaichen He Affiliation: School of Intelligence Science and Technology, Peking University Xinyi He Affiliation: School of Intelligence Science and Technology, Peking University Rui Huang Affiliation: University of Hong Kong Zilong Zheng Affiliation: State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China Zhenliang Zhang Affiliation: State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China Yiwu Zhong ⋆ Equal contribution † Corresponding author Affiliation: School of Intelligence Science and Technology, Peking University Affiliation: State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China" ], "source_url": "https://arxiv.org/html/2610.13021v1", "source_sha256": "144ba4b449d1ff386c07485bd0b5349298baf100174b7481b80495cf50f0b760", "source_bytes": 408405, "source_json_sha256": "f9a53f6b7c23e808512763104e296500cc404c3816c5a9771d6a1fa95d3a759a", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "TongSim构object-action任务/stategraph，basicexperience→advancedcompositionalunseen任务；offline交互选择含optimal/suboptimal/invalid可执行action候选的确定性benchmark。", "evaluation_evidence": "168basic+100advanced，ICL/skilllearning/agentharness对照；更多experience不持续提升，宽stepbudget也不替代经验。", "caveats": [ "offline candidateactionchoices不是自由物理控制部署", "optimal-path类别/advancedgoldtrajectory必须只在evaluator，oracleexperience/中间表示需独立diagnostic标签" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/NousResearch/hermes-agent", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.13021v1#S3", "https://arxiv.org/html/2610.13021v1#S4", "https://arxiv.org/html/2610.13021v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Benchmark Construction", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6404, 6804 ] ] }, { "section_index": 3, "heading": "4 Main Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4424, 4824 ] ] }, { "section_index": 4, "heading": "5 Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8271, 8671 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.13021.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12998", "assignment_index": 71, "title": "Evaluating LLM Creativity through Long-Tail Performance", "published": "2026-10-09T05:03:18Z", "authors": [ "Yichen Wang", "Tenghao Huang", "Vishakh Padmakumar", "Mina Lee" ], "affiliation": [ "Yichen Wang Affiliation: University of Chicago Email: yichenzw@uchicago.edu Tenghao Huang Affiliation: Microsoft Email: tenghaohuang@microsoft.com Vishakh Padmakumar Affiliation: Stanford University Email: vishakhp@stanford.edu Mina Lee Affiliation: University of Chicago Email: mnlee@uchicago.edu" ], "source_url": "https://arxiv.org/html/2610.12998v1", "source_sha256": "53815f132bc3b1eb5354a2d6a043431af78938b84de3329b9d2d7a43d5dc34e7", "source_bytes": 500461, "source_json_sha256": "f9bf956bde54824fa578e722baf6d08507d239bd44ed3fb7373daeac6f53dfda", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "固定reference输入输出分布，以Pile ngramabsence novelty选existingbenchmarktailfraction，q@Tail(alpha)为tail任务quality均值；对LM-Arena humancreativewritingrank做Kendall对照。", "evaluation_evidence": "15bench/26openmodels，tailness与rank相关；多数creativitypromoting方法提升peaknovelty却降低tailquality。", "caveats": [ "surface rarity不等语义/文化新颖，Pile不代表每个模型pretrain分布", "qTail非独立于通用能力/任务difficulty的creativityfactor；alpha/qualitychoice需独立validation" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/EQ-bench/creative-writing-bench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/JunyiYe/CreativeMath/blob/main/data/subset.json", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/X-PLUG/WritingBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/YichenZW/awesome_llm_creativity", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/YichenZW/longtail-creativity", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/allenai/MacGyver", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/anirudhlakkaraju/cs4_benchmark", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/lmarena/arena-hard-auto/blob/main/data/arena-hard-v2.0/question.jsonl", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/tatsu-lab/stanford_alpaca", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/QwQ-32B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-32B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-8B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3.5-27B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/allenai/OLMo-3-32B-Think", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/allenai/OLMo-3-7B-Instruct", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/allenai/OLMo-3.1-32B-Instruct", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/ChenLiu1996/Brainteaser", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/Pandores/aime-1983-2025", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/SAA-Lab/LitBench-Test-Release", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/euclaise/writingprompts", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/llm-aes/doc-storygen-v2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/roneneldan/TinyStories", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/swj0419/BookMIA", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/tatsu-lab/alpaca", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/tau/commonsense_qa", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-3-12b-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-3-27b-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-3-4b-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-3n-E4B-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-4-31B-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/microsoft/Phi-4", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-2506", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/nanbeige/Nanbeige4-3B-Thinking-2511", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-8B-v1", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-30B-A3B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/open-source/gpt-oss-20b", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12998v1#S3", "https://arxiv.org/html/2610.12998v1#S4", "https://arxiv.org/html/2610.12998v1#S6", "https://arxiv.org/html/2610.12998v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Evaluate LLM Creativity through Long-Tail Performance", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5033, 5433 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6148, 6548 ] ] }, { "section_index": 5, "heading": "6 Tailness Contributes to More Accurate Creativity Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8180, 8580 ] ] }, { "section_index": 9, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2098, 2498 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12998.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12981", "assignment_index": 72, "title": "The Evidence Anchoring Boundary: Where Visual Interventions Matter and When They Help in Multimodal LLMs", "published": "2026-10-09T04:26:50Z", "authors": [ "Qiji Zhou", "Junyi Hu", "Peiyu Wang", "Yue Zhang" ], "affiliation": [ "Qiji Zhou Affiliation: Westlake University Email: mailto:zhouqiji@westlake.edu.cnzhouqiji@westlake.edu.cn Junyi Hu Affiliation: Westlake University Affiliation: Duke University Email: mailto:junyi.hu@duke.edujunyi.hu@duke.edu Peiyu Wang Affiliation: Institute of Automation, Chinese Academy of Sciences Email: mailto:yue.zhang@wias.org.cnyue.zhang@wias.org.cn Yue Zhang Affiliation: Westlake University" ], "source_url": "https://arxiv.org/html/2610.12981v1", "source_sha256": "4e7681c86215ea3e9eb89efbc3269de67bc142256618a1d6aa994ad3fbdb0fec", "source_bytes": 383061, "source_json_sha256": "535c8c50d9c7a629c569b649bccd783dff5958ca5e8df18eb0e3264f9804a49e", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "depthmatchedRMS visual-residual edit定位operationalEAB；counterfactualsubtraction按presence/attribute/relation calibrationselective routing，保留原预测除非calibration支持该类型干预。", "evaluation_evidence": "六checkpoint三family，EAB frozen oncalibration+disjointeval；attribute五折image-disjoint routing，小提升而universal/跨calibratedscope可能harm。", "caveats": [ "EAB只对固定edit的leverage，不是所有visual信息普适锁死层", "effect≠correctness，goldbinarylabel仅calibration/scoring，cacheoracle pairs为diagnostic不直接部署" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12981v1#S3", "https://arxiv.org/html/2610.12981v1#S4", "https://arxiv.org/html/2610.12981v1#S5", "https://arxiv.org/html/2610.12981v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Task and Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5559, 5959 ] ] }, { "section_index": 3, "heading": "4 Q1: Where Can a Visual Intervention Matter?", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3563, 3963 ] ] }, { "section_index": 4, "heading": "5 Q2: When Should a Visual Intervention Be Used?", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7903, 8303 ] ] }, { "section_index": 5, "heading": "6 Evaluation: Benefits, Ablations, and Scope", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3760, 4160 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12981.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12961", "assignment_index": 73, "title": "MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph", "published": "2026-10-09T03:59:34Z", "authors": [ "Yingjie Zhu", "Han Cheng", "Xuefeng Bai", "Youcheng Pan", "Yang Xiang", "Kehai Chen" ], "affiliation": [ "Yingjie Zhu Affiliation: Harbin Institute of Technology, Shenzhen, China Affiliation: Peng Cheng Laboratory, Shenzhen, China {zhuyj,2024311560}@stu.hit.edu.cn , {xiangy,panych}@pcl.ac.cn , Email: chenkehai@hit.edu.cn Han Cheng Affiliation: Harbin Institute of Technology, Shenzhen, China Email: baixuefeng@hit.edu.cn Xuefeng Bai Affiliation: Harbin Institute of Technology, Shenzhen, China Youcheng Pan Affiliation: Peng Cheng Laboratory, Shenzhen, China {zhuyj,2024311560}@stu.hit.edu.cn , {xiangy,panych}@pcl.ac.cn , Yang Xiang Affiliation: Peng Cheng Laboratory, Shenzhen, China {zhuyj,2024311560}@stu.hit.edu.cn , {xiangy,panych}@pcl.ac.cn , Kehai Chen Affiliation: Harbin Institute of Technology, Shenzhen, China" ], "source_url": "https://arxiv.org/html/2610.12961v1", "source_sha256": "2a42bfe82aa3af3631073b9a0f8f739781f89f4f5c0d7781be86bcc3fb07d793", "source_bytes": 393454, "source_json_sha256": "4e0e4af8766f0dc638d5034622ee4798527fea4250d44d11840a7a006cf34d0a", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "MVGraph程序生成/验证partial/temporal多graphview确需多视图，真实渲染输入；8maingraphtasks+29stage perception/alignment/composition/reasoning诊断。", "evaluation_evidence": "8kmain+64kdiagnosticrecords，15LVLM；mainendtoend和progressivegoldintermediateoracle定位瓶颈，dropviews需保序temporal约束。", "caveats": [ "给goldintermediate的stageoracle只能diagnostic，不能宣称端到端视觉graph能力", "严格mainactor只见images/query，不能读latentgraph/answer/plan；datasetstage不同输入协议分开评分" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/AAAndy-Zhu/MVGraph", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12961v1#S3", "https://arxiv.org/html/2610.12961v1#S4", "https://arxiv.org/html/2610.12961v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 MVGraph : Benchmark Design and Construction", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6190, 6590 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1388, 1788 ] ] }, { "section_index": 4, "heading": "5 Results and Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12250, 12650 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12961.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12959", "assignment_index": 74, "title": "Video Codec-Guided Training-Free Token Compression for Video Large Language Models", "published": "2026-10-09T03:56:58Z", "authors": [ "Xihua Sheng", "Ruijie Liu", "Chang Wen Chen" ], "affiliation": [ "Xihua Sheng Ruijie Liu Chang Wen Chen Affiliation:" ], "source_url": "https://arxiv.org/html/2610.12959v1", "source_sha256": "60efeb55686c481409088738317fb82cef5a6fbf7e593a4bd29a16e4e0bb4ee3", "source_bytes": 325518, "source_json_sha256": "b8fa046e5ffc5a5a0dcd9eba63392333f12ab7fe394775969ad3fa27c582ea2a", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "真实codec referenceframe/motionvector/residual对齐token grid；residual只生成protection候选池，referencechainmotion只限定跨sampledframe merge区域，最终仍localvisualfeaturesimilarity决定。", "evaluation_evidence": "三VLLM四videobench对FlashVID，主要减少feature comparisons；residual相邻变化加权与阈值有消融。", "caveats": [ "不能用光流/随机motion替代压缩码流metadata核心", "comparisoncount下降不直接等于端到端walltime，需计codec解析与未被免除的visionencode成本" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12959v1#S3", "https://arxiv.org/html/2610.12959v1#S4", "https://arxiv.org/html/2610.12959v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "III Overview", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4987, 5387 ] ] }, { "section_index": 3, "heading": "IV Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13792, 14192 ] ] }, { "section_index": 4, "heading": "V Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12959.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12942", "assignment_index": 75, "title": "In-Context Learning Can Help Vision-Language Models Overcome Training Priors", "published": "2026-10-09T03:40:07Z", "authors": [ "Kun Wang", "Xindi Wu", "Sanghyuk Chun", "Olga Russakovsky", "Esin Tureci" ], "affiliation": [ "Kun Wang Xindi Wu Sanghyuk Chun Olga Russakovsky Esin Tureci Affiliation: Princeton University Affiliation: {bw5889, xw5676, sanghyukc, olgarus, eturci}@princeton.edu" ], "source_url": "https://arxiv.org/html/2610.12942v1", "source_sha256": "839729f5e226c755936005ff396f178f492a054fe7b3897114f08584b823494b", "source_bytes": 715763, "source_json_sha256": "3391ce6d405887c3e39bb5f63f62df9223577d3c4df7857b18ad65109b95b086", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "visualICL controlledprior-aligned/conflicting counterfactual任务，complex/simple/no-prior settings区分formatlearning、prioroverride与底层感知；同demonstration/promptnormalize跨模型。", "evaluation_evidence": "五visiontasks五VLMfamilies，机制attention/masking分析；highcardinalitycounting不稳定、部分10/15count低于zeroshot。", "caveats": [ "只说明某些可靠能力可被redirect，不证明fewshot能创建缺失能力", "demo可用训练label，querygold不可入prompt；counterfactual/images与类别选择要固定validation" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12942v1#S3", "https://arxiv.org/html/2610.12942v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 ICL Construction for Visual Perception and Reasoning", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7555, 7955 ] ] }, { "section_index": 3, "heading": "4 In-Context Learning Overcomes Training Priors", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14465, 14865 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12942.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12907", "assignment_index": 76, "title": "BLEND: Balancing Personalization vs. Generalization in Federated Vision-Language Models", "published": "2026-10-09T02:57:18Z", "authors": [ "Payam Abdisarabshali", "Fardis Nadimi", "Kasra Borazjani", "Naji Khosravan", "Richeng Jin", "Seyyedali Hosseinalipour" ], "affiliation": [ "Payam Abdisarabshali University at Buffalo–SUNY payamabd@buffalo.edu Fardis Nadimi University at Buffalo–SUNY Fardisna@buffalo.edu Kasra Borazjani University at Buffalo–SUNY Kasrabor@buffalo.edu Naji Khosravan Adobe naji.khosravan@gmail.com Richeng Jin Zhejiang University richengjin@zju.edu.cn Seyyedali Hosseinalipour University at Buffalo–SUNY alipour@buffalo.edu" ], "source_url": "https://arxiv.org/html/2610.12907v1", "source_sha256": "33f117ec6aaaad69392d054b843a165984badf4f1e3c56098d96ce15fa52f11a", "source_bytes": 1588735, "source_json_sha256": "f5466b01e1dde89f006e6a7a260b719e3c6ffd1efc9c7f83f523d4eaab0504c3", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "BLEND CLIP每layerglobalvision/textadapters跨clientaggregate，personalvisionprojection留local；blendglobal/personaloutputs并用frozenzeroshotprojectionheads anchor促进unseen泛化。", "evaluation_evidence": "10client simulatedFL、多rank/layer/γλsweep、3L40S；personalization/generalizationHmean tradeoff。", "caveats": [ "FL结构本身没有DP/secureagg privacy保证；serverless未验证", "仅textimage分类，audio/video/generative外推开放；γλ selection需要独立validation不能test调最佳" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/payamsiabd/BLEND.git", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12907v1#S3", "https://arxiv.org/html/2610.12907v1#S4", "https://arxiv.org/html/2610.12907v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "5 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 694 ], [ 294, 694 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12907.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12866", "assignment_index": 77, "title": "CaptchaAgent: Verifiable Visual Reasoning for CAPTCHA Solving", "published": "2026-10-09T01:50:26Z", "authors": [ "Shangquan Sun", "Ruoyu Chen", "Zhirui Liu", "Xiang Liu", "Chenghao Qian", "Wenqi Ren", "Xiaochun Cao" ], "affiliation": [ "Shangquan Sun Ruoyu Chen Zhirui Liu Xiang Liu Chenghao Qian Wenqi Ren Xiaochun Cao † † thanks: Shangquan Sun is with Nanyang Technological University, Singapore. † † thanks: Ruoyu Chen is with the School of Cyber Security, University of Chinese Academy of Sciences, China. † † thanks: Zhirui Liu, Wenqi Ren, and Xiaochun Cao are with the School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, No. 66, Gongchang Road, Guangming District, Shenzhen, Guangdong 518107, P.R. China. † † thanks: Xiang Liu is with Beihang University, Beijing, China. † † thanks: Chenghao Qian is with the University of Leeds, Leeds, U.K. † † thanks: This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible." ], "source_url": "https://arxiv.org/html/2610.12866v1", "source_sha256": "b12f862d37801643ffe11359c2d57c4db4a87ccd2f05b17623fde3b2c5bc1b93", "source_bytes": 575155, "source_json_sha256": "82c6f02f71ff5f163dfb3b8867b5f055464c8f3751058181a26dda352e587e32", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "研究structuredvisualstate与一致性诊断对有限视觉puzzle的作用；本审查只保留高层视觉推理/防御研究信息，不形成CAPTCHA自动破解实施方案。", "evaluation_evidence": "OpenCaptchaWorld20固定family browserexecutedresearchbenchmark与Oedipus小offlinejudge任务；后者非同一Pass1定义，支出只模型调用。", "caveats": [ "不纳入CAPTCHA自动破解/服务特定操作adapter实施范围；可另议不涉及访问控制的一般视觉puzzle诊断", "固定低熵family不证明production挑战泛化；稳定错误可通过consistency，ASCII丢连续几何" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/browser-use/browser-use", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/microsoft/playwright", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12866v1#S3", "https://arxiv.org/html/2610.12866v1#S4", "https://arxiv.org/html/2610.12866v1#S5", "https://arxiv.org/html/2610.12866v1#S6", "https://arxiv.org/html/2610.12866v1#S8" ], "evidence_sections": [ { "section_index": 2, "heading": "III Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10932, 11332 ] ] }, { "section_index": 3, "heading": "IV System Implementation Details", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2760, 3160 ] ] }, { "section_index": 4, "heading": "V Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6696, 7096 ] ] }, { "section_index": 5, "heading": "VI Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6563, 6963 ] ] }, { "section_index": 7, "heading": "VIII Limitations, Reproducibility, and Safety", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1754, 2154 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12866.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12864", "assignment_index": 78, "title": "Time Delay System for Sequence Modeling", "published": "2026-10-09T01:43:37Z", "authors": [ "Youshan Zhang", "Jialu Li" ], "affiliation": [ "Youshan Zhang † † thanks: This research was funded by the research\nproject of Chuzhou University (Grant No. 2025qd36). Affiliation: School of Artificial Intelligence, Chuzhou University, Chuzhou, China Email: youshanzhang@chzu.edu.cn Jialu Li † † thanks: Corresponding author. Affiliation: Department of Electrical and Computer Engineering, Boston University, Boston, USA Email: jli21@bu.edu" ], "source_url": "https://arxiv.org/html/2610.12864v1", "source_sha256": "3a30d2163a993f87a82fb9fd9d58329c172781fe698bd7aaa87d430045488caa", "source_bytes": 472173, "source_json_sha256": "e142deb71b1e3d85a500d0c68d794cf84ee46997a719f01afdeec2fde3e48f33", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "TDS semanticlagsoftattention取历史context/similarity，diagonalstate用closedformcumsum，TSKfuzzyprototype rules调instant与delayupdate；不是普通LSTMhiddenstate。", "evaluation_evidence": "IMDB分类、CNNDailyMail80ktrain+2kval/test文本生成，H100；仅两任务及lag数敏感性。", "caveats": [ "O1只可能针对固定lag/state更新，overall必须计历史attention窗口L和fuzzyrules成本", "官方split/subsampling/匹配baselinebudget尚需深入核，不能因headlineSOTA直接推广可扩展LLM" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/YoushanZhang/TDS", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12864v1#S3", "https://arxiv.org/html/2610.12864v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3238, 3638 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12864.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12783", "assignment_index": 79, "title": "Shift-Aware Identity-Guided Latent Refinement for Referring Audio--Visual Segmentation", "published": "2026-10-08T22:39:41Z", "authors": [ "Kun Li", "Sami Sebastian Brandt", "Michael Ying Yang" ], "affiliation": [ "Kun Li Affiliation: University of Liverpool, UK Sami Sebastian Brandt Affiliation: IT University of Copenhagen, DK Michael Ying Yang Affiliation: University of Bath, UK" ], "source_url": "https://arxiv.org/html/2610.12783v1", "source_sha256": "cd24a584bd27fdb09e655b05c9c5f50a4888f46f4c2821d28d623d17db7e2fa1", "source_bytes": 430068, "source_json_sha256": "855cad846b0879337892acd4d95d8d37057433a6fe7d297e77e79ac96d98859e", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SAIL tri-modalRefAVS的frame-dynamictoken与globalsemanticanchor分离，shiftawarelatentrefinement按domain证据纠漂移，SAM2promptdecoder+连续memory保持referentidentity。", "evaluation_evidence": "RefAVSBench officialvideo split含seen/unseen/null与R2AVSBench；目标变化/遮挡/声音衰弱是主要诊断情形。", "caveats": [ "需同步真实audio/video/query和learnedtokens/SAM2memory，图像classscore不代替tri-modal", "codeafterpublication未确认release，跨域/targetshift声明仍需完整protocol与independentvalidation" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12783v1#S3", "https://arxiv.org/html/2610.12783v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10192, 10592 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15154, 15554 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12783.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12734", "assignment_index": 80, "title": "StereoEngine: Scaling Stereo Bimanual Robot Data without Stereo Cameras", "published": "2026-10-08T21:28:58Z", "authors": [ "Jason Chen", "I-Chun Arthur Liu", "Gaurav Sukhatme", "Daniel Seita" ], "affiliation": [ "Jason Chen I-Chun (Arthur) Liu Gaurav S. Sukhatme Daniel Seita Affiliation: University of Southern California" ], "source_url": "https://arxiv.org/html/2610.12734v1", "source_sha256": "0698a5d5dc301554f655f3660efe5b456d139762d9188884f9af4c385d6935fe", "source_bytes": 236274, "source_json_sha256": "c015a999a667c7a0028ace67b02de016bca79ebb1da9853bc557af1e25a8e011", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "StereoEngineL monocularvideo depthwarp生成rightview后videodiffusion inpaintdisocclusion；S从simstereo联合生成photo-real左右，无realvideo输入；actualstereodata用于policy midtrain/finetune。", "evaluation_evidence": "两PiPERarms/ZEDMini2 realrollouts，3469公共monoepisodes各baseline同源转换；下游context实验仅proprio无languageinstruction。", "caveats": [ "必须真实diffusion/inpainting和双视图一致性，单depthwarptoy不算全机制", "realrobot数据与teleop依赖非本批资源；stereopolicy结果不自动证明通用VLA语言理解或productionAB" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/datasets/12e21/bi_piper_subset", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/Autobrik/bi_piper_pick_and_place_fruits_mantra", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/JohnnyXing/001_2026041415_bipiper", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/RDLwicked/pick_up_green_block_drawer_piperX", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/RoboSynChallenge/cobotmagic_Real_water_pouring", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/Sohaib03/piperx_pick_and_place_v1_2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/eval_piperx_pick_and_place_v1_4", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/eval_piperx_pick_and_place_v1_5", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/piperx_pick_and_place_v1", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/piperx_pick_and_place_v1_2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/piperx_pick_and_place_v1_3", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/csjmacmi/piperx_pick_and_place_v1_4", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/gauravpradeep/t01_piper_pick_and_place_bimanual_lerobot", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mjung11/bi_piper_open_tissue_test", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mjung11/eval_bi_piper_tissue_xvla", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mjung11/nc_bi_piper_folding_mj_A", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mjung11/nc_bi_piper_pot_car_apache_merged", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/mjung11/nc_bi_piper_pp_grape_pot_depth_mj_A", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/moonyh1230/piper_wooden_block_dual_realsense_0922", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/richardshkim/bi_piper_dataset", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12734v1#S3", "https://arxiv.org/html/2610.12734v1#S4", "https://arxiv.org/html/2610.12734v1#S5", "https://arxiv.org/html/2610.12734v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "III Problem Statement", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 991, 1391 ] ] }, { "section_index": 3, "heading": "IV Method: StereoEngine", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12878, 13278 ] ] }, { "section_index": 4, "heading": "V Real-World Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9092, 9492 ] ] }, { "section_index": 5, "heading": "VI Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 465, 865 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12734.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12711", "assignment_index": 81, "title": "Byzantine-Resilient Federated Large Language Model via Hypergraph Message Passing", "published": "2026-10-08T20:47:08Z", "authors": [ "Kai Li", "Hanlin Cai", "Zihao Liu", "Yichen Li", "Jose Alcaraz Calero", "Wei Ni", "Falko Dressler", "Özgür B. Akan" ], "affiliation": [ "Kai Li Hanlin Cai Zihao Liu Yichen Li Affiliation: Jose Alcaraz Calero, ,\nWei Ni, ,\nFalko Dressler, ,\nand Affiliation: Özgür B. Akan," ], "source_url": "https://arxiv.org/html/2610.12711v1", "source_sha256": "1fe8a1589810401a3c795fa6aa1718324c28136d87fbe9f5926f37e538737498", "source_bytes": 372038, "source_json_sha256": "5073a46e4005d66bfe99d3315a9ad56800521a2aff59723bc10ff61415588114", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "FedLLM frozenbase+LoRA clientupdates作为hypergraphnodes，HMP+GAE结合历史checkpoint关系/reconstruction检测和降权不一致更新；server无clientrawdata、benign身份未知。", "evaluation_evidence": "Qwen2.5/Llama3.2 AGNews/YahooAnswers分类，RTXPRO6000Blackwell；CSE/PPL与accuracy比较GAE及多defenses。", "caveats": [ "分类semanticentropy不等广泛factualtruth/knowledge安全保证", "需真实clienttraining/LoRAupdates/aggregation，已知攻击label不可给defense；无rawdata不构成DPprivacy保证" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/GuangLun2000/HMP-GAE", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12711v1#S3", "https://arxiv.org/html/2610.12711v1#S4", "https://arxiv.org/html/2610.12711v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "III Formulation of FedLLMs and Threat Model", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7986, 8386 ] ] }, { "section_index": 3, "heading": "IV HMP-GAE for Byzantine-Resilient FedLLMs", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "V Performance Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12711.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12703", "assignment_index": 82, "title": "Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation", "published": "2026-10-08T20:33:38Z", "authors": [ "Marco Stefani", "Edoardo Parpaiola", "Matteo Bino", "Gianluca Caregnato", "Francesco Roncolato", "Daniel Fusaro", "Loris Nanni", "Alberto Pretto" ], "affiliation": [ "Marco Stefani marcostefani.research@gmail.com Edoardo Parpaiola edoardo.parpaiola@gmail.com Matteo Bino matteo.bino.2002@gmail.com Gianluca Caregnato gianluca.caregnato@gmail.com Francesco Roncolato roncolatofra@gmail.com Daniel Fusaro daniel.fusaro@studenti.unipd.it Loris Nanni loris.nanni@unipd.it Alberto Pretto alberto.pretto@unipd.it organization=Department of Information Engineering (DEI), University of Padova,\ncity=Padova,\ncountry=Italy" ], "source_url": "https://arxiv.org/html/2610.12703v1", "source_sha256": "185652553eaef210de61b11e4a5d80e94b1538879af8858fdea14f4b7b49a5e0", "source_bytes": 640252, "source_json_sha256": "f5cfec8d5f2edfaa96f52416b7d28c7bd60f88f4ddedbb15a7aa09d57e7c9207", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "pretrainedResNet/ViT/ViM任务adapt后各自XAI图与human-gaze heatmap统一normalized spatialprobabilities，比JSS/CSS/Pearson，分析freeze/ensemble。", "evaluation_evidence": "fine-grained/medical/emotion三gazedata，Mamba出现near-unit supportoverlap却近零linearcorrelation。", "caveats": [ "human gaze alignment不是模型因果faithfulness或临床安全证明", "architecture-specificXAI不同存在方法混杂；原结果仅explainabilitydiagnostic，不能宣称实现更透明泛化能力" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/CaregnatoGianluca/VisionExplainability", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/LorisNanni/VisionExplainability", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/zhangyongshun/resnet_finetune_cub", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12703v1#S4", "https://arxiv.org/html/2610.12703v1#S5", "https://arxiv.org/html/2610.12703v1#S6" ], "evidence_sections": [ { "section_index": 3, "heading": "4 Datasets", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7787, 8187 ] ] }, { "section_index": 4, "heading": "5 Models", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5659, 6059 ] ] }, { "section_index": 5, "heading": "6 Experimental Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12703.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12650", "assignment_index": 83, "title": "Feature Competition and Its Representational Origins in Vision Language Models: Evidence from the Stroop Paradigm", "published": "2026-10-08T19:13:43Z", "authors": [ "Nurbüke Teker", "Rui Xiao", "Zeynep Akata", "Shuchen Wu" ], "affiliation": [ "Nurbüke Teker Rui Xiao Zeynep Akata Shuchen Wu" ], "source_url": "https://arxiv.org/html/2610.12650v1", "source_sha256": "1e375ee521f2d59ae9e54495bf5ce1f8bfa11756fa0b653a94a8677d8df75101", "source_bytes": 322140, "source_json_sha256": "259a8e084f69d81f2e25db9d0a2aa6bcf4a1fd48795d0998d2be6101db9042f5", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Stroop100 word×ink stimuli比较contrastive/generativeVLM行为，ink-only/word-only/combinedembedding RSA与conceptchunk centroidsteering分离文字/颜色feature影响。", "evaluation_evidence": "10congruent90incongruent，embedintervention仅CLIP；wordsteering较易、inkdirection弱且generative输出分类有edgecases。", "caveats": [ "color可能更distributed，centroid弱steering不证明颜色功能绝对弱", "synthetic100图和CLIPintervention不可推广全部现代VLM；operator/goldconceptaxes仅diagnostic" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12650v1#Sx3", "https://arxiv.org/html/2610.12650v1#Sx4", "https://arxiv.org/html/2610.12650v1#Sx5", "https://arxiv.org/html/2610.12650v1#Sx8" ], "evidence_sections": [ { "section_index": 2, "heading": "Evaluating VLMs on the Stroop Task", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12128, 12528 ] ] }, { "section_index": 3, "heading": "The Representation of Word Shape and Ink Color in the Image Embedding Space", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2225, 2625 ] ] }, { "section_index": 4, "heading": "Steering the Concept Representations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3359, 3759 ] ] }, { "section_index": 7, "heading": "Discussion and Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1752, 2152 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12650.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12575", "assignment_index": 85, "title": "Mind the Temporal Gap: Evaluating Video Models for Temporal Understanding in Driving", "published": "2026-10-08T18:02:22Z", "authors": [ "Shiva Sreeram", "Erfan Aasi", "Tsun-Hsuan Wang", "Alaa Maalouf", "Daniela Rus" ], "affiliation": [ "Shiva Sreeram Erfan Aasi Tsun-Hsuan Wang Alaa Maalouf and Daniela Rus MIT CSAIL University of Haifa † † thanks: *Research was sponsored by the Department of the Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-19-2-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Department of the Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Author Alaa Maalouf thanks the Neubauer Family Foundation and the Council for Higher Education’s MAOF program for their support. $ˆ1$MIT CSAIL, $ˆ2$University of Haifa ." ], "source_url": "https://arxiv.org/html/2610.12575v1", "source_sha256": "24cbdd35b322d185a135019a40ad473f0f119f204f4ff319838dacbea0cb7285", "source_bytes": 227778, "source_json_sha256": "73dcafc743dbbb19a36a266411a8cdec12464ca5560aa41fd46800db001a7168", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "egomotion multi-levelprobe比较standard/visual/fewshot/CoT，LoRA和视觉特征压缩前/后1layerCNN probe定位信息缺失或语言readout失败。", "evaluation_evidence": "CARLA controlledsim与小realdriving仅standardprompt，较新模型fewshot有taskgain但类biaspersist；learning/probe结果限Sim。", "caveats": [ "strongmotionprobe不等于原VideoLLM安全驾驶/worldmodel能力", "real样本/任务范围小，prompt依赖不一致；对安全关键部署不能给泛化保证" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12575v1#S2", "https://arxiv.org/html/2610.12575v1#S3", "https://arxiv.org/html/2610.12575v1#S4" ], "evidence_sections": [ { "section_index": 1, "heading": "II Probing via Different Levels.", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11089, 11489 ] ] }, { "section_index": 2, "heading": "III Experimental Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "IV Conclusion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2214, 2614 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12575.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12461", "assignment_index": 86, "title": "OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs", "published": "2026-10-08T17:59:46Z", "authors": [ "You-Zhe Xie", "Ting-Wei Chou", "Yu-Hsuan Li", "Kaipeng Zhang", "Zhixiang Wang", "Yu-Lun Liu" ], "affiliation": [ "You-Zhe Xie Affiliation: National Yang Ming Chiao Tung University Affiliation: Alaya Lab Ting-Wei Chou Affiliation: National Yang Ming Chiao Tung University Yu-Hsuan Li Affiliation: National Yang Ming Chiao Tung University Kaipeng Zhang Affiliation: Alaya Lab Zhixiang Wang Affiliation: Alaya Lab Corresponding authors Yu-Lun Liu Affiliation: National Yang Ming Chiao Tung University Corresponding authors" ], "source_url": "https://arxiv.org/html/2610.12461v1", "source_sha256": "1e697597bea79be0b8a3803246d215c0651dfdf482e375ad5796339b77af9d61", "source_bytes": 437116, "source_json_sha256": "07d97062c6ff6106a27923889b997036fc435f8aec52532a4f2b6203c46c2315", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "OuroWorld用VLM选referenceview/推dynamic→真实videogeneration，再VGGTlift/TrajectoryCrafter补多view；Fourierperiodic4DGS deformation与reference-groundedDriftField吸收viewinconsistency。", "evaluation_evidence": "39 reconstructed/generated3DGS scenes，GTfreeautomaticmetrics+humanpairwise；去periodiccycle seam坏而overallmotion指标不明显变。", "caveats": [ "生成prior不合理motion会传播，宽views更disocclusion", "loop数学periodicity不是物理dynamics正确；本质完整生成pipeline，静态3DGS+手写sinwave只是conceptdemo" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12461v1#S3", "https://arxiv.org/html/2610.12461v1#S4", "https://arxiv.org/html/2610.12461v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8763, 9163 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13290, 13690 ] ] }, { "section_index": 4, "heading": "5 Conclusion and Limitation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 588, 988 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12461.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12458", "assignment_index": 87, "title": "OmniCapBench: A Deep-Structured Evaluation Framework for Fine-Grained Audio-Visual Captioning", "published": "2026-10-08T17:59:21Z", "authors": [ "Zhongyu Yang", "Jiale Tao", "Ruitao Chen", "Zuhao Yang", "Yingfang Yuan", "Xueliang Zhao", "Auden", "Kai Wang", "Shuai Shao", "Biao Wang", "Steve Yves", "Qinglin Lu" ], "affiliation": [ "Zhongyu Yang ♥ Affiliation: Hunyuan, Tencent Jiale Tao ♥ Affiliation: Hunyuan, Tencent Ruitao Chen ♥ Affiliation: Hunyuan, Tencent Zuhao Yang Affiliation: Nanyang Technological University Yingfang Yuan Affiliation: Northumbria UniversityProject Website: https://01yzzyu.github.io/OmniCapBench/ Xueliang Zhao Affiliation: Hunyuan, Tencent Auden Affiliation: Hunyuan, Tencent Kai Wang Affiliation: Hunyuan, Tencent Shuai Shao Affiliation: Hunyuan, Tencent Biao Wang ✉ Affiliation: Hunyuan, Tencent Steve Yves ✉ Affiliation: Hunyuan, Tencent Qinglin Lu Affiliation: Hunyuan, Tencent" ], "source_url": "https://arxiv.org/html/2610.12458v1", "source_sha256": "d2ad8894e91f5707e79165f43f645125538db971e99b9fbca5904c623d911936", "source_bytes": 1654817, "source_json_sha256": "53530968960c1fad6e1c0196618d190f9e9a384a40e8e999ffb01382c3d34684", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "OmniCapBench model直接从真实audio-video生成entityreference/shot/audioevent三trackatomicJSONunits；先deterministicstructural约束再localizedsemanticLLMcompare，避免wholecaption单分数。", "evaluation_evidence": "786densevideos多omnimodel；specializedfreeformcaptioners因structuredinstructionpoor被排除主表，identitydrift/crossmodalalignment是难点。", "caveats": [ "格式instructionfollowing仍可能混入perception分数，不能把排除模型比成视觉能力差", "LLM semanticjudge仍是proxy，referenceunit/goldtiming不入native-generationactor；claimguarantee不能替代human校准" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Jingkang50/FunQA", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/MME-Benchmarks/Video-MME", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/QwenLM/Qwen2.5-Omni", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/NJU-LINK/OmniVideoBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/longvideobench/LongVideoBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/omni-research/Tarsier2-Recap-585K", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/zai-org/VisionRewardDB-Video", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/openbmb/MiniCPM-o-2_6", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12458v1#S3", "https://arxiv.org/html/2610.12458v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 OmniCapBench Benchmark", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13725, 14125 ], [ 6118, 7218 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9875, 10275 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12458.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12451", "assignment_index": 88, "title": "VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation", "published": "2026-10-08T17:58:47Z", "authors": [ "Boyao Han", "Chen Shi", "Jingjing Qian", "ZhuoTan Tian", "Li Jiang" ], "affiliation": [ "Boyao Han  Chen Shi  Jingjing Qian  Zhuotao Tian  Li Jiang The Chinese University of Hong Kong, Shenzhen Shenzhen Loop Area Institute  Harbin Institute of Technology, Shenzhen Project Page: https://boyaohan.github.io/VersaCamVLA.github.io/" ], "source_url": "https://arxiv.org/html/2610.12451v1", "source_sha256": "3ccedc1057a0bb84732928d9f4d373d522e7e3962c42350685727d06b2e988a2", "source_bytes": 463110, "source_json_sha256": "7ca9e0e291caa07131aa7e776dfc30ef2c17efe58d00d1f3ccec11e7e9f0a91c", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "VersaCamVLA把variableposedRGB集合成fixedscenetokens，stageI RGB/semantic/edge targetview预测+WristAugmentedPoseSampling；stageII丢decoder、frozenceneencoder+jointspatialcompressor补pretrainedVLA视觉条件。", "evaluation_evidence": "RoboTwin2/LIBERO和realrobot camera-count/unseenpose，combinesemantic+edge比RGBonly更好。", "caveats": [ "仍需knowncamera pose/extrinsics，不是任意未知camera即插即用", "真实policyactionrollout与scene-tokentraining必需；没有infer3D传感/render不等没有几何metadata" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12451v1#S3", "https://arxiv.org/html/2610.12451v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7577, 7977 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12451.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12427", "assignment_index": 89, "title": "FastBench: Can Streaming VLMs Perceive High-Dynamic Real-World Streams?", "published": "2026-10-08T17:55:20Z", "authors": [ "Yuxuan Hu", "Weikang Shi", "Yang Bo", "Xudong Lu", "Xintong Guo", "Shuhan Li", "Yuyang He", "Huankang Guan", "Peiwen Sun", "Yunqiao Yang", "Wenbo Li", "Rui Liu", "Hongsheng Li" ], "affiliation": [ "Yuxuan Hu † † thanks: Equal contribution. Affiliation: CUHK MMLab Weikang Shi Affiliation: CUHK MMLab Yang Bo Affiliation: Huawei Research Xudong Lu Affiliation: CUHK MMLab Xintong Guo Affiliation: Huawei Research Shuhan LI Affiliation: Huawei Research Yuyang He Affiliation: Huawei Research Huankang Guan Affiliation: Huawei Research Peiwen Sun Affiliation: CUHK MMLab Yunqiao Yang Affiliation: CUHK MMLab Wenbo Li Affiliation: CUHK MMLab Rui Liu Affiliation: Huawei Research Hongsheng Li Affiliation: CUHK MMLab" ], "source_url": "https://arxiv.org/html/2610.12427v1", "source_sha256": "f561945a7ffb1fc8cf52dffb5cd54ee89098999411b1574be0f2dd675db50887", "source_bytes": 251480, "source_json_sha256": "2a83c42bbb897095c28675e8c01e0658b55e5c5bccc72641876b652779ac42b7", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "FastBench高FPStrajectory-groundedQA过滤2FPS可答样本、人核evidenceinterval；ProactiveFrame在1secstreamchunk通过texttoken调整未来FPS，近期高FPSFocusWindow+较旧sparseGeneralWindow。", "evaluation_evidence": "306QA、forward/instant/backward三时序scope与多domain，denserframes24FPS后饱和/30FPS下降；oracleintervalfocusing明显强于model自主rate。", "caveats": [ "human evidenceinterval只external/oracle诊断，实际actor不能预看未来frames/gold事件时段", "dual-tier固定context有detail-historytradeoff；不能只用全video离线答题假称streaming或去除无响应失败" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Ashone3/FastBench", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12427v1#S3", "https://arxiv.org/html/2610.12427v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 FastBench", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11593, 11993 ], [ 8651, 9751 ] ] }, { "section_index": 3, "heading": "4 Experiment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9361, 9761 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12427.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12402", "assignment_index": 91, "title": "SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models", "published": "2026-10-08T17:44:17Z", "authors": [ "Hongxing Li", "Jinyue Su", "Dingming Li", "Wenqi Zhang", "Weiming Lu", "Jun Xiao", "Yueting Zhuang", "Yongliang Shen" ], "affiliation": [ "Hongxing Li † † thanks: Equal contribution. † Corresponding author. Affiliation: Zhejiang University Email: hongxing.li@zju.edu.cnhttps://github.com/ZJU-REAL/SpaceCast-Bench-0.15em[height=0.95em]figures/github-logo.pdfGitHubhttps://huggingface.co/datasets/hongxingli/SpaceCast-Bench-0.15em[height=0.95em]figures/hf-logo.pdfHuggingFace Jinyue Su 1 1 footnotemark: 1 Affiliation: Zhejiang University Email: syl@zju.edu.cnhttps://github.com/ZJU-REAL/SpaceCast-Bench-0.15em[height=0.95em]figures/github-logo.pdfGitHubhttps://huggingface.co/datasets/hongxingli/SpaceCast-Bench-0.15em[height=0.95em]figures/hf-logo.pdfHuggingFace Dingming Li Affiliation: Zhejiang University Wenqi Zhang Affiliation: Zhejiang University Weiming Lu Affiliation: Zhejiang University Jun Xiao Affiliation: Zhejiang University Yueting Zhuang Affiliation: Zhejiang University Yongliang Shen 2 2 footnotemark: 2 Affiliation: Zhejiang University" ], "source_url": "https://arxiv.org/html/2610.12402v1", "source_sha256": "9073196d483567b87899fd3b6a0195860683a1f4404fdcadb8da71cccbfbbefc", "source_bytes": 560093, "source_json_sha256": "7cab5e345e5c2dd9d4343c4d58a4e66e5996e6ee96dc0282f53c833dea30d4ca", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SpaceCast observe-transform-infer只提供变换前orderedRGBviews，自然语言T之后模型推未观察S1；programmaticverification分static/local/global预测及桥接views控制。", "evaluation_evidence": "182realScanNet/+scenes3862Q、21models，syntheticSFT/GRPO和sixOOD；explicit3D evidence比生成outcome媒介更稳。", "caveats": [ "3Dprivileged-input比较不得宣称RGBonly；actor不能看futureS1渲染/gtstate", "programmatictrain与benchmarkscene/source隔离需完整核，不能从in-domain升幅推universalsimulation能力" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "mailto:hongxing.li@zju.edu.cnhttps://github.com/ZJU-REAL/SpaceCast-Bench-0.15em[height=0.95em]figures/github-logo.pdfGitHubhttps://huggingface.co/datasets/hongxingli/SpaceCast-Bench-0.15em[height=0.95em]figures/hf-logo.pdfHuggingFace", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "mailto:syl@zju.edu.cnhttps://github.com/ZJU-REAL/SpaceCast-Bench-0.15em[height=0.95em]figures/github-logo.pdfGitHubhttps://huggingface.co/datasets/hongxingli/SpaceCast-Bench-0.15em[height=0.95em]figures/hf-logo.pdfHuggingFace", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12402v1#S3", "https://arxiv.org/html/2610.12402v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 SpaceCast-Bench", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6871, 7271 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12402.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12390", "assignment_index": 92, "title": "Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search", "published": "2026-10-08T17:39:11Z", "authors": [ "Ziming Dai", "Dabiao Ma", "Ziheng Guo", "Jack Dong", "Zimu Zhou" ], "affiliation": [ "Ziming Dai Note: Both authors contributed equally to this research. Affiliation: City University of Hong Kong , Hong Kong , China email: phoenix.dai@my.cityu.edu.hk , Dabiao Ma Affiliation: Qfin Holdings, Inc. , Beijing , China email: madabiao-jk@qifu.com , Ziheng Guo Affiliation: Tianjin University , Tianjin , China email: sygzh6@tju.edu.cn , Jack Dong Affiliation: Carnegie Mellon University , Pittsburgh , USA email: jack.dong@bytedance.com and Zimu Zhou Affiliation: City University of Hong Kong , Hong Kong , China email: zimuzhou@cityu.edu.hk" ], "source_url": "https://arxiv.org/html/2610.12390v1", "source_sha256": "6d8a45d94d93f77faf6a0516a886047cd2d50497d47e61bc5f422dc56e7be372", "source_bytes": 243771, "source_json_sha256": "8b25912c06a2f1776e33270c52de65a79313a004b3b244c42d97c2c3c1bbea19", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "qualified industrial production/explicit online A/B found in full text; NEXT BATCH, not frozen16", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "LLMBlockFE offline只appendimmutablefeaturecodeblock，真实execute/downstreamAUCgate+localBayesianparamtuning；depthcalibratedcreditrollback，多independenttrajectory只共享fixed方向描述，freezebestprograms供online不调用LLM。", "evaluation_evidence": "CFPBTimely/GossipCop+两private，五生产riskapps2025.9–11同样本parallelKS比较2694342用户；原文另报告onlineAB约RMB1B放款、审批数不变、baddebt8.0%→7.7%。", "caveats": [ "生产数字为作者报告，AB段未披露group人数/randomization/CI，不能算本地独立复现", "私有riskdata不能冒充公共；下一批实际programsearch仅用公开data+独立validation/test和sandbox" ], "online_ab": "Explicitly reported real-world online A/B in §5.3: approx RMB1B disbursements, unchanged approved-applicant count, bad-debt rate 8.0% to 7.7%; statistical/randomization details not established by the reviewed passage.", "upstream_assets": [ { "url": "https://github.com/ByteDance-Seed/seed-oss", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12390v1#S3", "https://arxiv.org/html/2610.12390v1#S4", "https://arxiv.org/html/2610.12390v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3. Notations and Problem Formulation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4868, 5268 ] ] }, { "section_index": 3, "heading": "4. Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13727, 14127 ] ] }, { "section_index": 4, "heading": "5. Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14519, 14919 ], [ 6147, 9147 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12390.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12355", "assignment_index": 94, "title": "Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models", "published": "2026-10-08T17:22:24Z", "authors": [ "Hongxing Li", "Yixin Li", "Dingming Li", "Zixuan Wang", "Yuchen Yan", "Wenqi Zhang", "Weiming Lu", "Yongliang Shen" ], "affiliation": [ "Hongxing Li Yixin Li Dingming Li Zixuan Wang Yuchen Yan Wenqi Zhang Weiming Lu Yongliang Shen \\corresponding" ], "source_url": "https://arxiv.org/html/2610.12355v1", "source_sha256": "c3e7a85930bb55c1d3a373b2488d2ebb323227856a8db456fb1874acdd6e3153", "source_bytes": 168770, "source_json_sha256": "dd4356c83a40b7492623a055e32c4aaf56cb88909124d9b128165c6093356fce", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "GPD offline从3Dscan生成depth/semantic/BEV compacttext且questionconditionroute给同模型teacher+trainreferenceanswer；GRPO加privilegedKL仅incorrectstudenttrajectories，deployRGBonly。", "evaluation_evidence": "五spatialbench samebackbone/dataGRPO/OPSD controlled对照，3D+answerprivilege互补、只wrongKL与routing消融。", "caveats": [ "geometry/referenceanswers严格仅trainingteacher，evalstudent不能看到gt3D答案", "必须真实on-policyrollout+wrongtrajectorymask/KL，teacherfewshot直接推理不是蒸馏复现" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12355v1#S3", "https://arxiv.org/html/2610.12355v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10529, 10929 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7957, 8357 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12355.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12338", "assignment_index": 95, "title": "VFold: Symmetry-Aware Cross-Layer Value Cache Compression", "published": "2026-10-08T17:14:28Z", "authors": [ "Neha Verma", "Sungwon Kim", "Kenton Murray", "Kevin Duh" ], "affiliation": [ "Neha Verma Affiliation: Johns Hopkins University Email: nverma7@jhu.edu Sungwon Kim Affiliation: Johns Hopkins University Email: kmurra23@gmu.edu Kenton Murray Affiliation: George Mason University Email: kevinduh@cs.jhu.edu Kevin Duh Affiliation: Johns Hopkins University" ], "source_url": "https://arxiv.org/html/2610.12338v1", "source_sha256": "8f2f7694c5a6f8d55d59d33e75d21092cdbcf5eaf87fdaa0809ce694d7e2cc8d", "source_bytes": 452897, "source_json_sha256": "018b340950714207b423351077a372bfdba081224cc6f84b0b965b8220b072f2", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "VFold CCAcalibration得adjacentvaluebasisalignment并fold WV/WO functionpreservingweights，runtimevalues用后merge到sharedrunningmean跨layers；keys/RoPE保持不合并。", "evaluation_evidence": "Llama8B/Mistral24B/Qwen8B LongBench/RULER，50%value减仅25%totalKV；还有quant/keyprune组合。", "caveats": [ "fullvalue当前token可计算不意味着历史mergedcache或decodeoutput完全无误差", "真实cachealias/share及weightfold是核心，离线CKA/均值MSE模拟不算；key不能套同symmetry" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/nverma1/vfold", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12338v1#S3", "https://arxiv.org/html/2610.12338v1#S4", "https://arxiv.org/html/2610.12338v1#S5", "https://arxiv.org/html/2610.12338v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Preliminary analysis and motivation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5596, 5996 ] ] }, { "section_index": 3, "heading": "4 VFold", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5542, 5942 ] ] }, { "section_index": 4, "heading": "5 Experimental Settings", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 5, "heading": "6 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12338.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12382", "assignment_index": 93, "title": "WorldAlign: Decoupled 4D Reward for World-Consistent Video Generation", "published": "2026-10-08T17:35:19Z", "authors": [ "Jing He", "Kaixin Ding", "Xingye Tian", "Guibao Shen", "Wenhang Ge", "Xin Tao", "Pengfei Wan", "Ying-Cong Chen" ], "affiliation": [ "HKUST (Guangzhou)/HKUST; KlingAI; HKU (official PDF p1)" ], "source_url": "https://arxiv.org/pdf/2610.12382v1", "source_sha256": "32a7dc1682308baad3e18d3653c10fc11c15ff7d2c7fabbbdd67472970c4b479", "source_bytes": 8745380, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "official PDF fallback; KlingAI practical generation", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "official PDF pages1..7 first3000 extracted characters each; not complete15pages", "defining_method": "语义subjectmasktracker而非motion检测分静/动；GFM staticpointcloud maskedreprojection+camera-motion hinge防staticcollapse；frozenVLM sample-specificdynamic/physics/shape/texturebinarychecklistreward，samepromptgroup各rewardnormalize后GRPO。", "evaluation_evidence": "DAVIS/COCO1389samples，100randomheldout；Wan2.1/Wan2.2 matcheddata/settings，rewardindependentmaskedepipolar及trackedDINO/vbench检查。", "caveats": [ "onlinepost-training不是productionuserAB，VLMjudge不等物理正确groundtruth", "WorldR1 reproduction省camera-awarewarp/trajectoryreward，VGGRPO因无officialcode排除；baseline并非原完整系统", "真实tracker/GFM/VLM+I2Vpolicyupdates全部核心，imageclassscore/reprojectiontoy不可替代" ], "online_ab": "online refers to group-rollout policy training; no quantified production-traffic A/B established in reviewed PDF passages", "upstream_assets": [ { "url": "https://worldalign.github.io/", "status": "official project page declared; asset contents not independently verified" } ], "evidence_anchors": [ "https://arxiv.org/pdf/2610.12382v1#page=5", "https://arxiv.org/pdf/2610.12382v1#page=6", "https://arxiv.org/pdf/2610.12382v1#page=7" ], "evidence_sections": [ { "format": "PDF", "pages": [ 1, 2, 3, 4, 5, 6, 7 ], "first_extracted_characters_per_page": 3000 } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12382.pdf", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12327", "assignment_index": 96, "title": "SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference", "published": "2026-10-08T17:06:32Z", "authors": [ "Qitong Wang", "Xinwei Niu", "Mingluo Su", "Shanwei Zhao", "Shiai Zhu", "Huan Wang" ], "affiliation": [ "Qitong Wang  \nXinwei Niu  \nMingluo Su  \nShanwei Zhao  \nShiai Zhu  \nHuan Wang Affiliation: Westlake University Affiliation: Ant Group https://wang-qitong.github.io/SparseDecoding" ], "source_url": "https://arxiv.org/html/2610.12327v1", "source_sha256": "7a670983a55e4b6bc7f4c2843a586a59773b4365edef7545d72b369053a536b8", "source_bytes": 430429, "source_json_sha256": "4628b86b23560ac8c6341c8400f4f3534b3d9ed95187c2bfa703ded09b1e121d", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SparseDecoding真实dense ARgeneration收layeractivations且排除prefill形成Hessian/calibration矩阵，再OBS/SparseGPT/Wanda剪枝；N:MSpMV custombitmask索引fixed-step遍历针对decodebatch1。", "evaluation_evidence": "Llama8/70B与Qwen14/32B长文/codegen质量，A100nativeSpMV end-to-enddecode测量与C4cal对照。", "caveats": [ "必须selfgenerateddecode activations不能naturaltext/randomfixture替代", "置零后densematmul不是kernel加速，SparseSpMM也不能冒称SpMV；实际source/kernel不可用则只mechanismdiag" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12327v1#S3", "https://arxiv.org/html/2610.12327v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Proposed Method: SparseDecoding", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10519, 10919 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12346, 12746 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12327.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12310", "assignment_index": 97, "title": "Is In-Domain Training Enough for Fine-Grained Industrial Anomaly Understanding?", "published": "2026-10-08T16:59:29Z", "authors": [ "Xingwu Zhang", "Duanyang Du", "Huiling Zhu", "Jiayue Dai", "Yixiao Liu", "Guozhi Liu", "Zhihan Zhang", "Zijun Long" ], "affiliation": [ "Xingwu Zhang * , Duanyang Du * , Huiling Zhu * , Jiayue Dai , Yixiao Liu , Guozhi Liu , Zhihan Zhang , Zijun Long Hunan University \nUniversity of Aberdeen \nSouth China Normal University South China University of Technology" ], "source_url": "https://arxiv.org/html/2610.12310v1", "source_sha256": "988229c4188a39ba3e189f84a70f929aebc6fa1c94f2953c0c42699a3420ea15", "source_bytes": 408475, "source_json_sha256": "f77722f031461854712661258d198bfb8b90d24d46ffb965a3aef086d6079324", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SiGMA normalreference/industrialknowledgesearch+heterogeneousMLLMreaders+真实visualdefectexpert，query-ref evidencecalibration；unlabeledpredictions估taskcompetence/queryevidence质量后reliability聚合。", "evaluation_evidence": "MMAD七task，trainedspecialist定位仍弱/detection可差于base；agent-only互补不能补fineperception，expert reader有call/cost增加。", "caveats": [ "label-freecontroller仍需实际predictions/evidence，不可testgold估可靠性", "three-reader route在AppendixD选，validation隔离待更深核；benchmark非真实生产factoryAB" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12310v1#S3", "https://arxiv.org/html/2610.12310v1#S4", "https://arxiv.org/html/2610.12310v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 3, "heading": "4 Experiments and Analysis", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10014, 10414 ] ] }, { "section_index": 4, "heading": "5 Conclusion and Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1482, 1882 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12310.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12305", "assignment_index": 98, "title": "From What to Which: Decoding Modifier Grounding in Frozen MLLMs", "published": "2026-10-08T16:53:59Z", "authors": [ "Barbara Toniella Corradini", "Caterina Gallegati", "Ludovica Genovese", "Vittorio Murino" ], "affiliation": [ "Barbara Toniella Corradini Affiliation: AI for Good (AIGO), Istituto Italiano di Tecnologia, Italy Email: barbara.corradini@iit.it Caterina Gallegati Affiliation: University of Siena, Italy Email: c.gallegati@student.unisi.it Ludovica Genovese Affiliation: AI for Good (AIGO), Istituto Italiano di Tecnologia, Italy Affiliation: University of Genoa, Italy Email: ludovica.genovese@iit.it Vittorio Murino Affiliation: AI for Good (AIGO), Istituto Italiano di Tecnologia, Italy Affiliation: University of Verona, Italy Email: vittorio.murino@iit.it" ], "source_url": "https://arxiv.org/html/2610.12305v1", "source_sha256": "8284731acc1558aba5b5d7f0514d64fcf3b2f67085a29afdfa8599cb26300ccc", "source_bytes": 315867, "source_json_sha256": "00acf0d302e9e6a352136dbf52164c1700cdbc970bbd087f23d041783a7110ae", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "OTTER frozenMLLMgenerated-token/visualpatchstates监督投影，semantic/attention/spatialcost和learneddustbin/marginals做entropyregularizedSinkhornOT，compactpatchmap+optionalresidual。", "evaluation_evidence": "RefCOCO controlledmodifier/noun、perturb/freegeneration/RefCOCOgtransfer；transfer用fixedfullOT而主表adaptivegated，backbonelayerselection另appendix。", "caveats": [ "supervisedprobe可解码不等base模型内生因果解释", "跨数据换headoperatingpoint不能无条件称同一个头泛化；selectedregion/gtmask仅外部supervision/scoring" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12305v1#S3", "https://arxiv.org/html/2610.12305v1#S4", "https://arxiv.org/html/2610.12305v1#S5", "https://arxiv.org/html/2610.12305v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Evaluation Protocol and Controls", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5211, 5611 ] ] }, { "section_index": 3, "heading": "4 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8642, 9042 ] ] }, { "section_index": 4, "heading": "5 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3612, 4012 ] ] }, { "section_index": 5, "heading": "6 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9786, 10186 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12305.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12266", "assignment_index": 99, "title": "DVD: Dynamic Vector Decoding for Efficient MLLM-based Perception", "published": "2026-10-08T16:32:28Z", "authors": [ "Jinghua Hou", "Zhe Liu", "Hengshuang Zhao" ], "affiliation": [ "Jinghua Hou Zhe Liu † † thanks: Project leader. Hengshuang Zhao † † thanks: Corresponding author. Affiliation: The University of Hong Kong Affiliation: https://almoonysl.github.io/projects/DVD" ], "source_url": "https://arxiv.org/html/2610.12266v1", "source_sha256": "794087b6cf7fadeadc59577053ac1b03872999ef1ca32b3896efda7e8ab00616", "source_bytes": 351278, "source_json_sha256": "a74ee2a7991264c3dd18473e54dd6b315395195085ea3418cb280cb245008323", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "DVD将2Dbox/mask/3Dbox编码1Dvectors，经learned1Dencoderdecoder+统一VQcodebook变少量perceptiontokens，真实MLLM输出这些tok再detokenize连续geometry。", "evaluation_evidence": "RefCOCO系列和SUNRGBD/KITTI/Hypersim/nuScenes，2D/3Djointperception与token/latency；需所有codec/MLLM训练阶段。", "caveats": [ "有限learnedcodebook不保证任意unboundedrange/precisionOOD", "直接4floathead或固定coordbins不是同机制；必须实际vocab/codeprediction与detokenizer并计encode成本" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12266v1#S3", "https://arxiv.org/html/2610.12266v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8519, 8919 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12266.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12242", "assignment_index": 100, "title": "TokenRouter: Efficient Serving System for Token-Level LLM Routing", "published": "2026-10-08T16:21:50Z", "authors": [ "Tianyu Fu", "Tengxuan Liu", "Ruoxi Wang", "Yixin Dong", "Yi Ge", "Yichen You", "Yu Wang" ], "affiliation": [ "Tianyu Fu † † thanks: Equal contribution. Affiliation: Tsinghua University Tengxuan Liu Ruoxi Wang Affiliation: Tsinghua University Yixin Dong Affiliation: Carnegie Mellon University Yi Ge,\nYichen You,\nYu Wang † † thanks: Corresponding author: Yu Wang (yu-wang@tsinghua.edu.cn). Affiliation: Tsinghua University" ], "source_url": "https://arxiv.org/html/2610.12242v1", "source_sha256": "ce243cfd9db0d74a91ecf2b6239dfd887158f004ecbe3ddee1545bda70157ae2", "source_bytes": 349961, "source_json_sha256": "c74ebe6a6822c985e9d09c80988a3dd2da24f37a7603bb999818f9b6d377e522", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "TokenRouter requestcentric route/send/receive coroutine，modelcentric subservers异步tri-loop/handoffresume，延后batchadmission阈值由DTMC吞吐模型据N/routingprob/step latency确定。", "evaluation_evidence": "SGLangbackend五tokenrouting算法，Qwen0.6/8/32B、多pairs/workloads；全部并发、perrequest outputtokens/elapsed吞吐定义。", "caveats": [ "真实subservers/modelstatehandoff/cache和scheduler必需，离散queue simulator不是系统复现", "throughput定义/负载/routeprob/等待latencytradeoff不能外推所有productionQoS；不是质量改进算法" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/thu-nics/TokenRouter", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12242v1#S3", "https://arxiv.org/html/2610.12242v1#S4", "https://arxiv.org/html/2610.12242v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Programming Interface", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6285, 6685 ] ] }, { "section_index": 3, "heading": "4 System Design", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6649, 7049 ] ] }, { "section_index": 4, "heading": "5 Experiment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7312, 7712 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12242.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12147", "assignment_index": 101, "title": "Healthy Counterfactual Generation via Diffusion Inpainting for Mammography Classification", "published": "2026-10-08T15:31:56Z", "authors": [ "Inês Cruchinho Garcia", "Mariana Mourão", "Francisco Maria Calisto", "Carlos Santiago", "Jacinto Nascimento" ], "affiliation": [ "Inês Cruchinho Garcia Mariana Mourão Francisco Maria Calisto Carlos Santiago Jacinto Nascimento Affiliation: Institute for Systems and Robotics, Instituto Superior Técnico, Lisbon, Portugal E-mail ines.cruchinho.garcia@tecnico.ulisboa.pt" ], "source_url": "https://arxiv.org/html/2610.12147v1", "source_sha256": "2f5918f138b62b8d45ccb3177331859db70eeaf003c4d05f4b0f5f0ad74e4f1c", "source_bytes": 119161, "source_json_sha256": "229f1b46ea38a176dc670a0332efe2bb4c0f728c0cd07bb2ad68371a5aafde45", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "在健康 BI-RADS 1 乳腺图上训练 DDPM，再用 RePaint 在已标注病灶框内重绘正常组织；保留框外结构，并把生成图作为负类训练增广，而非训练通用基础模型。", "evaluation_evidence": "VinDr-Mammo 5000 检查/20000 图像，官方16000训练、4000测试；512缩放，训练集增加1410健康反事实。ConvNext、ViT、MammoCLIP、FPN-MIL敏感度多数改善，但特异度下降且部分ROC-AUC下降；另报告固定80%特异度下召回及放射科医师真实性判断。", "caveats": [ "本研究主要是乳腺影像任务增广，非当前foundation-model方法范围，故不新增本批实现。", "未见配对预测一致性验证；病灶框内正常重绘及FID/FRD并不建立因果病灶移除或临床收益。", "保留敏感度/特异度权衡与负结果，不能仅挑改善项。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/ines03garcia/diffusion-based-counterfactual-generation", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12147v1#S2", "https://arxiv.org/html/2610.12147v1#S3", "https://arxiv.org/html/2610.12147v1#S4" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2012, 2412 ] ] }, { "section_index": 2, "heading": "3 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3423, 3823 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7890, 8290 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12147.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12144", "assignment_index": 102, "title": "Language-Specific Effects of Tokenizer Choice in Multilingual Language Models", "published": "2026-10-08T15:29:37Z", "authors": [ "Clara Meister", "Gül Sena Altıntaş", "Antoine Bosselut" ], "affiliation": [ "Clara Meister Email: clara.meister@epfl.ch Gül Sena Altıntaş Affiliation: EPFL, University of Toronto, Vector Institute Antoine Bosselut" ], "source_url": "https://arxiv.org/html/2610.12144v1", "source_sha256": "3b10a00bd6ef20bdc3259a6151ed6724533bc3c80cdaec5944f588b3744a0ea3", "source_bytes": 725196, "source_json_sha256": "0f0d184f5388ef8c84f86d65b0aeba13c3565bdb6eae130d92a7ba9eb615450e", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "以同一nanochat d24 1.27B架构、训练语料/配方/词表约128K比较54种tokenizer、123个LM；按语言分析BPB对tokenizer配置的敏感性，并做单因素配置和tokenizer训练语言配比重分配。", "evaluation_evidence": "31训练语言，28有词边界；每语言BPB变化与LM训练份额负相关。英语独占tokenizer训练损害几乎全部语言，equal/inverse配比不必改善低资源语言，尤其LM训练重复数据时。XNLI/MultiBLiMP有区分，XCOPA/Belebele在该规模近随机。", "caveats": [ "54-tokenizer面板不是随机试验；跨配置指标关联不自动构成因果。", "多数tokenizer单训练seed，只有5种做重复seed；family-cluster稳健误差不能替代全部模型重跑。", "文字系统/资源量相关，重复训练交互明显；结论限单架构、单配方、31语言与主要语言建模指标。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/karpathy/nanochat", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/HuggingFaceTB/finemath", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12144v1#S2", "https://arxiv.org/html/2610.12144v1#S3", "https://arxiv.org/html/2610.12144v1#S4", "https://arxiv.org/html/2610.12144v1#S5", "https://arxiv.org/html/2610.12144v1#S8" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8882, 9282 ] ] }, { "section_index": 2, "heading": "3 How much does tokenizer choice matter across languages? (RQ1)", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3514, 3914 ] ] }, { "section_index": 3, "heading": "4 Which Languages Are Most Tokenizer-Sensitive? (RQ2)", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5184, 5584 ] ] }, { "section_index": 4, "heading": "5 Do design choices change per-language outcomes? (RQ3)", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8177, 8577 ] ] }, { "section_index": 7, "heading": "8 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1191, 1591 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12144.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12133", "assignment_index": 103, "title": "Rehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read", "published": "2026-10-08T15:22:50Z", "authors": [ "Zhiyun Shi" ], "affiliation": [ "Zhiyun Shi Affiliation: Nanyang Technological University , Singapore email: ZHIYUN004@e.ntu.edu.sg" ], "source_url": "https://arxiv.org/html/2610.12133v1", "source_sha256": "ef293f6b2b9913c4557855dfd101748786badeb41dedfda98df87db0600ba287", "source_bytes": 631636, "source_json_sha256": "ba8c900f34f7a216aa13b3420eb6a7456d5f62866fa387132743077ee6b47eed", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Attic-KV在未知未来query时，用模型自身生成的、引用上下文内容的QA及内容自适应anchor token来rehearsal后压缩KV；作为KVzip/KVgrad/RestoreKV+的rehearsal替换，不把未来测试query送入缓存选择。", "evaluation_evidence": "Qwen3-4B/8B/14B及Llama系列，RULER和LongBench；比较不同keep ratio、固定比例rehearsal与compress-once answer-many，并报告压缩成本。650道RULER开发问题与全部测试隔离。改善因类别/host方法而异，少数multi-doc QA项持平或稍差。", "caveats": [ "只rehearse真正needle或将来读取跨度的oracle实验是机制诊断，不能作为可部署查询无关方法。", "下一批需复现真实self-generated QA/anchor及开发集调参和独立测试；不能用gold query/答案代替生成。", "收益需连同self-QA生成/压缩开销评估，不能仅比较解码后指标。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/NVIDIA/kvpress", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12133v1#S3", "https://arxiv.org/html/2610.12133v1#S4", "https://arxiv.org/html/2610.12133v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3. A Cache Keeps What It Rehearses", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6631, 7031 ] ] }, { "section_index": 3, "heading": "4. Two Principles and Attic", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8729, 9129 ] ] }, { "section_index": 4, "heading": "5. Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12483, 12883 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12133.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12126", "assignment_index": 104, "title": "SuperNav: An Agentic Navigation System for Any Task in Any Scene", "published": "2026-10-08T15:19:40Z", "authors": [ "Jinkai Zhang", "Jingyi Xu", "Yuanhong Yu", "Jiarui Guo", "Ruizhen Hu", "Hujun Bao", "Xiaowei Zhou", "Sida Peng" ], "affiliation": [ "Jinkai Zhang Affiliation: Zhejiang University Jingyi Xu Affiliation: Zhejiang University Yuanhong Yu Affiliation: Zhejiang University Jiarui Guo Affiliation: Zhejiang University Ruizhen Hu Affiliation: Shenzhen University Hujun Bao Affiliation: Zhejiang University Xiaowei Zhou Affiliation: Zhejiang University Sida Peng Affiliation: Zhejiang University Affiliation: Causa Robotics † Corresponding author" ], "source_url": "https://arxiv.org/html/2610.12126v1", "source_sha256": "8f3f220afabf303cf5d0e5b44c2446b4524f38457c2cd62e2b41e21b96eaf037", "source_bytes": 269005, "source_json_sha256": "16ef9535f5327d819fcb9eb9d319b4e80cd5744778ee59f89286361e62ea802b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SuperNav以冻结MLLM为导航规划器，通过Skills说明、像素点视觉接口和真实导航工具构成harness；执行器提供几何/学习式底层控制，任务进度与上下文循环管理高层计划。", "evaluation_evidence": "Habitat-GS、AI2THOR、HM3D比较GPT-5.6 Terra加不同执行器；另有Unitree Go2四视角RGB+LiDAR实机对象寻找演示。接口与历史输入和原生baseline不完全一致，实机演示不是生产A/B。", "caveats": [ "Skills文本不等于程序化约束，结果属于完整harness与执行器组合而非单独模型能力。", "下一批仅模拟器诊断可常规复现；物理机器人、安全操作需新授权及实际资源。", "输入/历史配置差异必须列出，实机少量对象任务不建立普遍导航可靠性。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12126v1#S3", "https://arxiv.org/html/2610.12126v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Navigation Agent Harness", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8601, 9001 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6670, 7070 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12126.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12060", "assignment_index": 105, "title": "Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning", "published": "2026-10-08T14:42:57Z", "authors": [ "Yicheng Xue", "Han Wu", "Jufeng Yang", "Minjing Dong", "Xinghao Chen", "Hanting Chen", "Jianyuan Guo" ], "affiliation": [ "Yicheng Xue Affiliation: City University of Hong Kong Affiliation: Zhejiang University Email: yichengxue@zju.edu.cn Han Wu Affiliation: Peking University, Email: jianyguo@cityu.edu.hk Jufeng Yang Affiliation: Nankai University Minjing Dong Affiliation: City University of Hong Kong Xinghao Chen Affiliation: Huawei Technologies Hanting Chen Affiliation: Huawei Technologies Jianyuan Guo" ], "source_url": "https://arxiv.org/html/2610.12060v1", "source_sha256": "fbbd09f31ae69938478851cee41da685f68208ddf5712f09dac7c8951aed4d6f", "source_bytes": 595139, "source_json_sha256": "b177af95a574db39676cb50cd75232a0f1a1b2d41a29ff1aa718fee3808b5ad1", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ViMoD在冻结Qwen3-VL-4B中保留全部原始Fine视觉/text KV，DART生成内容自适应Coarse groups及容量；TRACE随解码历史执行SELECT/HOLD/REPLACE Fine group。教师预告将需区域仅用于训练，学生轨迹用RLOO正确性与成功条件下occupancy目标。", "evaluation_evidence": "20%目标活动视觉预算，8个bench，新增训练参数约0.0546%；报告活动访问/总耗时下降。峰值内存42.67GiB高于Full的38.40GiB且跨预算基本不变，因为完整Fine/Coarse bank一直保留。", "caveats": [ "这不是KV整体内存压缩；不能用访问token预算改善宣称峰值显存改善。", "Coarse初始化和动态路由有开销，短输出或持续广域视觉读取收益可能不足。", "固定路由间隔、漏选group与teacher监督限制证据访问；跨backbone/解码配置鲁棒性未验证。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.12060v1#S3", "https://arxiv.org/html/2610.12060v1#S4", "https://arxiv.org/html/2610.12060v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9148, 9548 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15366, 15766 ] ] }, { "section_index": 4, "heading": "5 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 616 ], [ 216, 616 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12060.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.12048", "assignment_index": 106, "title": "Do Not Train Away Uncertainty: Early Uncertainty Anchored Calibration", "published": "2026-10-08T14:34:26Z", "authors": [ "Yutong Xie", "Jiawei Tang", "Zhenglin Hua", "Yuxiang Ma", "Si Qin", "Yaxin Hou", "Hui Liu", "Junhui Hou", "Yuheng Jia" ], "affiliation": [ "Yutong Xie , Jiawei Tang , Zhenglin Hua , Yuxiang Ma , Si Qin , Yaxin Hou , Hui Liu , Junhui Hou , Yuheng Jia † † thanks: Corresponding author. Email: {xyt,jw_tang,220252358,220256453,siqin,yaxin,yhjia}@seu.edu.cn Email: h2liu@sfu.edu.hk, jh.hou@cityu.edu.hk Affiliation: Key Laboratory of New Generation Artificial Intelligence Technology and ItsInterdisciplinary Applications (Southeast University), Ministry of Education, China Affiliation: School of Computing Information Sciences, Saint Francis University, Hong Kong, China Affiliation: School of Software Engineering, Southeast University, Nanjing, China Affiliation: School of Computer Science and Engineering, Southeast University, Nanjing, China Affiliation: Department of Computer Science, City University of Hong Kong, Hong Kong, China" ], "source_url": "https://arxiv.org/html/2610.12048v1", "source_sha256": "bd0b8b91ffa6e09d3b83e1be328828e40f6f6f03dbf70de4105505f0a6ac94a7", "source_bytes": 562016, "source_json_sha256": "6cbbe18246c7208aa8f7176441e05ee58a2ba8e14b70526c6ebf8c79f4c0ae76", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "EUA-Cal冻结早期checkpoint作为不确定性锚点；EPR对早期预测和当前预测做KL，PSR用早期真实类置信度加权归一化特征构造固定class prototype，再将余弦相似度经temperature变为soft target、对当前预测做KL，与CE联合训练。", "evaluation_evidence": "CIFAR10/100、TinyImageNet多种CNN/ViT；SciQ、ARC-E/C、CSQA上Llama3/3.1-8B与Qwen3-8B-Base LoRA。比较ACC/ECE/AECE与OOD AUROC/FPR95及校准baselines，报告tau变化分析。", "caveats": [ "早期checkpoint、alpha/beta等选择必须仅用训练/验证，不能按测试ECE挑锚点。", "作者建议prototype温度按训练NLL设定；下一批仍需隔离验证与测试，复现两项KL而非单独temperature scaling。", "精度和校准是不同目标，不把有限分类/MCQA证据推广成生成式不确定性保证。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.12048v1#S3", "https://arxiv.org/html/2610.12048v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11051, 11451 ], [ 1500, 4500 ], [ 4400, 7200 ], [ 7100, 9900 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 650, 2100 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.12048.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11993", "assignment_index": 107, "title": "DataVista: Diagnosing Multimodal LLMs on Data Video Understanding", "published": "2026-10-08T14:02:35Z", "authors": [ "Yupeng Xie", "Zhenyang Wang", "Jiayi Zhu", "Yinghao Tang", "Zhouan Shen", "Yiyu Chen", "Yuyu Luo" ], "affiliation": [ "Yupeng Xie Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Zhenyang Wang Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Jiayi Zhu Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Yinghao Tang Affiliation: Zhejiang University Zhouan Shen Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Yiyu Chen Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Yuyu Luo † † thanks: Corresponding author: Yuyu Luo (E-mail: yuyuluo@hkust-gz.edu.cn) Affiliation: The Hong Kong University of Science and Technology (Guangzhou)" ], "source_url": "https://arxiv.org/html/2610.11993v1", "source_sha256": "c3a977f1c133915c94f4179d5a9a99e8cbec97f58d58470d69894a1beed29e88", "source_bytes": 552322, "source_json_sha256": "1b02ca3dc11648d07197803c6cf113b7606a4134800201620ef51f78eb2958cc", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "DataVista从真实数据叙事视频构造三级data perception/temporal reasoning/narrative understanding及10类问题；人机协作问题生成、文本过滤后专家回视频核验/修改歧义，并保留多选精确集合评分。", "evaluation_evidence": "961视频、6775问题、19MLLM；主评统一50均匀帧无音频，有/无subtitle同一帧，overall合并两条件。单选exact match、多选必须选项集合完全正确。最佳Gemini3.1Pro70%，另诊断帧数/subtitle及chart reading、causal/narrative难项。", "caveats": [ "人类看完整视频并听原音频，而模型主表50帧无音频，故人类差距不是严格同信息条件能力差。", "视频URL、许可/可得性、重复视频分组、判分格式需下一批独立确认。", "原文称首次benchmark不自动代表通用视频理解覆盖；nextbatch为真实视频评估器，不做新模型claim。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/HKUSTDial/DataVista", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11993v1#S3", "https://arxiv.org/html/2610.11993v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 DataVista : Design and Construction", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9351, 9751 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12567, 12967 ], [ 600, 2400 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11993.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11954", "assignment_index": 109, "title": "Look Where You Can: Active View Selection for CAD Reconstruction under Occlusion", "published": "2026-10-08T13:40:38Z", "authors": [ "Kartik Bali", "Mahish Guru", "Yiderigun Borjigin", "Alexandra Starostina", "Christian J. Cyron", "Roland Aydin" ], "affiliation": [ "Kartik Bali Affiliation: Helmholtz-Zentrum Hereon Email: kartik.bali@hereon.de Mahish K. Guru Affiliation: Leuphana Universität Lüneburg Yiderigun Borjigin Affiliation: Universität des Saarlandes Alexandra Starostina Affiliation: Helmholtz-Zentrum Hereon Christian J. Cyron Affiliation: Technische Universität Hamburg Roland Aydin Affiliation: Universität des Saarlandes" ], "source_url": "https://arxiv.org/html/2610.11954v1", "source_sha256": "688d78a7fbbb48215e1667dc9cf17ee0c58eae2ccb3252ed6dc1631a2e7a7472", "source_bytes": 398419, "source_json_sha256": "37cea1816eeb4437e7e89aee1a7365ef40c314e2699670d8f43688684c2c6cd1", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SightCAD以冻结DINOv2图像特征+pose编码+cross-view transformer从完整可行候选view集选择K；warm start蒸馏SA-max与VLM多视角CadQuery SFT；随后selector用生成/执行后voxelIoU的GRPO、generator在组内最高reward views做GT-code SFT；最后固定selector独立generator GRPO，避免两policy信用混淆。", "evaluation_evidence": "CAD-Recode/Text2CAD约4k heldout，冻结相同场景placement/occlusion；DeepCAD/Fusion360及T-LESS、MP6D synthetic/real。报告invalid code、valid-onlyChamfer/mIoU及invalid记0的effective mIoU；viewpolicy按固定metric尺度，cross-method各自unitcube归一化。", "caveats": [ "训练奖励/visibility gate使用GT geometry仅训练可用；推理planner不可读GT CAD。", "selector读取全部候选render，因此K预算仅约束generator输入，不等价真实相机仅拍K次的主动感知成本。", "cross-method尺度归一化不评绝对尺寸/placement；真实图test-timecanonicalization与无预处理系统不可混称同协议。", "下一批需真实CadQuery执行与几何reward，不能用字符串/模板代码评分替代。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11954v1#S3", "https://arxiv.org/html/2610.11954v1#S4", "https://arxiv.org/html/2610.11954v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Dataset", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2411, 2811 ] ] }, { "section_index": 3, "heading": "4 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9169, 9569 ], [ 600, 3100 ], [ 4600, 8500 ] ] }, { "section_index": 4, "heading": "5 Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11920, 12320 ], [ 0, 1800 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11954.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11948", "assignment_index": 110, "title": "When History Helps and Hurts: Selective History Use across Multimodal Turns", "published": "2026-10-08T13:36:12Z", "authors": [ "Shuoyang Sun", "Kerui Gu", "Hao Fang", "Shaoli Huang", "Bin Chen" ], "affiliation": [ "Shuoyang Sun Kerui Gu 1 1 footnotemark: 1 Hao Fang Shaoli Huang Bin Chen Harbin Institute of Technology, Shenzhen \nAgiBot Tsinghua Shenzhen International Graduate School, Tsinghua University † † thanks: Equal contribution. † † thanks: Project lead. † † thanks: Corresponding author." ], "source_url": "https://arxiv.org/html/2610.11948v1", "source_sha256": "fcd027e5be73654a76c0bf7b380e30d27798ca451576bf36acf4995e11ddfe74", "source_bytes": 605844, "source_json_sha256": "48f7d9724d2e90f7f26adb84d6535ed5e03787668a7245487914fc98cacd5e5f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ReTurn将任务携带历史Reconfirm/Reground与证据携带历史Retrieve/Rebind做成保持target question/media/answer的配对局部对照，实际模型生成历史；配套restatement、基于显式reference的attention reweighting及Short OpenQA SFT/answer-preserving JS penalty。", "evaluation_evidence": "7000base任务、视觉/音频、13模型matched singleturn/multiturn；median OpenQA93.7%到72.3%。Qwen3Omni单seed DevTrain509pairs/1018views、204updates；SFT+3.42pp OpenQA但LongMCQ Reconfirm-1.3pp，Inv仅+.02/.20pp且interval跨0。", "caveats": [ "Reconfirm历史答案恰正确，不能独立证明查看当前媒体；题型配对保持答案但需要检查derivation-root/media-source隔离。", "显式reference reweighting只适用于该benchmark结构，不默认是通用自然会话media resolver。", "SFT新错90题/修复261题，Inv没有明确增益，须保留负结果；仅单模型单seed。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/collections/XiaomiMiMo/mimo-v25", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/HuggingFaceFV/finevideo", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11948v1#S3", "https://arxiv.org/html/2610.11948v1#S4", "https://arxiv.org/html/2610.11948v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 ReTurn: Benchmarking Selective History Use", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4346, 4746 ], [ 650, 2600 ] ] }, { "section_index": 3, "heading": "4 Evaluating Selective History Use", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11238, 11638 ] ] }, { "section_index": 5, "heading": "6 Mitigating Selective History Failures", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2650, 3050 ], [ 0, 1300 ], [ 1200, 3050 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11948.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11986", "assignment_index": 108, "title": "FearCaut-Qwen: Affective Steering in a Vision-Language Model Shifts the Decision Criterion for Hazard Assessment", "published": "2026-10-08T13:57:46Z", "authors": [ "Xiaoshan Zhou" ], "code_links": null, "affiliation": [ "The University of Sydney, School of Project Management (official PDF p1)" ], "source_url": "https://arxiv.org/pdf/2610.11986v1", "source_sha256": "8df0a7c6aeafa1b3033802c88c1ce6ebdbcc5cfc54e1705e8737ddee761b91e1", "source_bytes": 19722750, "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual official PDF fallback; mechanistic VLM diagnostic", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "official PDF selected page prefixes (explicit ranges below), not complete48pages", "defining_method": "在独立FindingEmo discovery上末prompt-token residual逐层linear probe、validation选L22-24；计算fear-vs-neutral classmean residual direction，validation选非退化alpha、内容hash冻结后在Qwen2.5VL-7B两阶段SeisMLLM perception/retrieval/posting中注入。另SwiGLU sparse neuron knockout用matched controls验证源情绪因果必要性，和distributed方向分别检验迁移。", "evaluation_evidence": "官方65建筑test不变，train-fit1039/validation202；emotion844discovery/280validation/280test ID互斥。fear alpha+.5 Red recall .270→.757但d'1.521→1.028、criterion+1.354→-.161；perception-only重现94.4%效应，sparse neurons target迁移无效。配对case bootstrap10000次及多重比较校正。", "caveats": [ "这是criterion lever而非更好damage discrimination；false Red也增加，部署代价选择未验证，不能宣称工程安全收益。", "sadness distributed direction全部退化，无法完成同为负valence控制；实际fear/sadness arousal未按计划分离。", "groundtruth stage1只用于train-fit structural-circuit诊断，不能在target评估推理输入提供gold；retrieval bank只train-fit。", "65cases单backbone，macroF1多重比较只有1/24存活，baseline阈值调整可得相近macroF1变化；不实现真实建筑安全决策。" ], "online_ab": "controlled offline intervention on65public cases, not production A/B", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/pdf/2610.11986v1#page=7", "https://arxiv.org/pdf/2610.11986v1#page=16", "https://arxiv.org/pdf/2610.11986v1#page=20", "https://arxiv.org/pdf/2610.11986v1#page=23", "https://arxiv.org/pdf/2610.11986v1#page=29", "https://arxiv.org/pdf/2610.11986v1#page=32", "https://arxiv.org/pdf/2610.11986v1#page=33" ], "evidence_sections": [ { "format": "PDF", "pages": [ 1, 2, 3, 4, 5, 6, 7, 8 ], "first_extracted_characters_per_page": 3000 }, { "format": "PDF", "pages": [ 9, 10, 11, 12, 13, 14, 15, 16, 17 ], "first_extracted_characters_per_page": 2200 }, { "format": "PDF", "pages": [ 21, 22, 28, 29, 30, 31, 35 ], "first_extracted_characters_per_page": 2000 }, { "format": "PDF", "pages": [ 20, 23, 32, 33 ], "first_extracted_characters_per_page": 2600 } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11986.pdf", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11938", "assignment_index": 111, "title": "Does Target Alignment Mean Target Recovery? An Evidence-Ladder Study of Adversarial Claims on Contrastive Encoders", "published": "2026-10-08T13:30:52Z", "authors": [ "Tao Yang", "Jianying Zhou" ], "affiliation": [ "Tao Yang Affiliation: Singapore University of Technology and Design Email: tao_yang@mymail.sutd.edu.sg Jianying Zhou Affiliation: Singapore University of Technology and Design Email: jianying_zhou@sutd.edu.sg" ], "source_url": "https://arxiv.org/html/2610.11938v1", "source_sha256": "5d77aa12091435d49aa409f3703ba6838f586d3328f26482dd45eb4d50376a05", "source_bytes": 248032, "source_json_sha256": "b40e0bd904d768dbd67244d143d7cc4663f3e1feb5c120031fb8db13b054fcac", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "将contrastive encoder目标攻击证据分成L0intent/L1victimcosinealignment/L2独立模型recovery/L3downstream/L4human。用训练/架构不同的supervised ResNet50、补充ViTjudge，真实target图blend正控、shuffledtarget负控；目标logitmargin变化减16个shuffle变化为ITE，预冻结50%blend reference后confirmatory匹配攻击。", "evaluation_evidence": "六encoder、三扰动预算，robust FARE/TeCoA/PMG/TRADES独立证据高于vanilla CLIP/SigLIP，但无cell达到blend-derived reference，前三在replication band故实际recovery仍未定。robust内部alignment/ITErho .24-.51，vanilla CLIP近0，跨encoder无单调关系。", "caveats": [ "50%真实图blend只是可复现reference，不是utility/人类语义阈值；没有L3/L4证据。", "robust训练预算与测试4/16/64除255不同，需报告固定checkpoint/hash/预算匹配与classcluster bootstrap。", "只考虑本地授权公开图像的防御评测，不扩大为外部目标或绕过系统。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11938v1#S3", "https://arxiv.org/html/2610.11938v1#S4", "https://arxiv.org/html/2610.11938v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 An Evidence Taxonomy for Adversarial Claims", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2049, 2449 ] ] }, { "section_index": 3, "heading": "4 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6130, 6530 ], [ 700, 2800 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11937, 12337 ], [ 0, 1300 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11938.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11924", "assignment_index": 112, "title": "Revisiting Identity and Spectra Dispersion in Media-Bridged Time Series Forecasting: Linking Multivariate Signals and Narrative Flows", "published": "2026-10-08T13:20:49Z", "authors": [ "Jierui Lei", "Wenjian Zhang", "Qingyi Yang", "Yuyang Hong", "Fangzheng Chen", "Zhengbo Zhang", "Haina Tang", "Shiming Xiang" ], "affiliation": [ "Jierui Lei Wenjian Zhang Qingyi Yang Yuyang Hong Fangzheng Chen Zhengbo Zhang Affiliation: Haina Tang, Shiming Xiang" ], "source_url": "https://arxiv.org/html/2610.11924v1", "source_sha256": "677550be2e6a925e92fe1d5fa8f4fe69d0db7028665ab6d2e19fc01e0811fe44", "source_bytes": 1229866, "source_json_sha256": "12fd308b554e20a06fdabf71e100d3f8d8545c3b5a55890eae179b5a8e25db21", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "MIDAPN先以frozen PLM将预对齐叙事变为数值pseudo-variates，再DFF频域融合、MIDAG static essence/dynamic behavior/latentcommonality及CIM图聚合。SPConv对FFT实/虚系数做局部卷积、层级avgpool压缩与通道扩张，按temporal reconstructioncapacity约束对整数expansion进行mixed-order搜索。", "evaluation_evidence": "13multivariate和12预对齐numerical-text数据；16native TSF与14TSFM/fusedPLM，TaTS统一prealignment，官方描述ETT/PEMS6:2:2、其他7:1:2，MAE/MSE；另fixedlongcontext比较不同预训练TSFM。", "caveats": [ "media明确不是原始图像/音频/视频，仅temporally aligned variables和PLM textpseudo-variates；不能当通用原始multimodalfoundation模型。", "SPConv容量成本是结构proxy、实际有效rank训练前未知；理论intallocation不是性能最优保证。", "作者代码链接的真实内容/可用性尚未独立核验；下一批需时间顺序split和新闻时间可用性防泄漏。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/leijieruilq/MIDAPN/tree/main", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11924v1#S3", "https://arxiv.org/html/2610.11924v1#S4", "https://arxiv.org/html/2610.11924v1#S5", "https://arxiv.org/html/2610.11924v1#S14" ], "evidence_sections": [ { "section_index": 2, "heading": "III Preliminaries and Media Pre-alignment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2679, 3079 ] ] }, { "section_index": 3, "heading": "IV Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "V Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8882, 9282 ], [ 0, 1700 ], [ 1700, 3500 ] ] }, { "section_index": 14, "heading": "XIV Clarifications for Media Scope", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 454, 854 ], [ 0, 854 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11924.html", "source_is_research_input_not_authority": true, "evidence_original_html_passages": [ { "section_id": "S4.SS4", "heading": "IV-D Search-Guided Spectral Prism", "read_character_ranges_in_bs4_section_text": [ [ 0, 2400 ], [ 5000, 7200 ] ], "source_url": "https://arxiv.org/html/2610.11924v1#S4.SS4" } ] }, { "arxiv_id": "2610.11907", "assignment_index": 113, "title": "Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations", "published": "2026-10-08T13:08:44Z", "authors": [ "Shuran Ma", "JiaLe Li", "Yuxin Dong", "Shan Zheng", "Qingyun Jiang", "Xiang Chen", "Qi Zhu", "Deyi Ji", "Yifan Yang", "Jianfeng Pan", "Yu Tian", "Xue Yang" ], "affiliation": [ "Shuran Ma JiaLe Li Yuxin Dong Shan Zheng Qingyun Jiang Xiang Chen Qi Zhu Deyi Ji Yifan Yang Jianfeng Pan Yu Tian Xue Yang Affiliation: [ Affiliation: [ Affiliation: [ Affiliation: [ Affiliation: [ Affiliation: [" ], "source_url": "https://arxiv.org/html/2610.11907v1", "source_sha256": "63e6c30d708c464964876289717f42ffebaf63ef77802b3012fcab08954a96d4", "source_bytes": 385857, "source_json_sha256": "87a7c8a6438730f8e030949e6f0830e069895ef4e513f334bc4e9bfae076e151", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "AIMS在native per-head query/key空间对视觉与prefill文本keycache均值prototype、生成context用最新生成token key作为summary；((cos(q,p)+1)/2)^(1/tau_source)给head/timestep/source自适应权重，目标s为加权prototype和，qtilde=(1-alpha)q+alpha*s，然后原attention/KV不变、无辅助forward。", "evaluation_evidence": "LLaVA1.5-7B/Qwen2.5VL-3B等，CHAIR500COCO初步分析与greedy/beam/nucleus主比较，多source ablation；报告CHAIRs/i、objectrecall/F1、captionlength及一般能力。CHAIR减少有部分recall/length下降，nucleus部分项收益较小。", "caveats": [ "训练free不等零开销，query hook与prototype更新应真实计时。", "同一500COCO初步发现与后续main选择可能存在选择偏差，下一批需专门validation与sourcegroup独立test。", "不能仅按短caption降低hallucination比例宣称内容质量提升；保留recall和长度变化。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/VisionXLab/AIMS", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/VisionXLab/AIMS_Benchmarks", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11907v1#S3", "https://arxiv.org/html/2610.11907v1#S4", "https://arxiv.org/html/2610.11907v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Preliminary Study", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3284, 3684 ] ] }, { "section_index": 3, "heading": "4 method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5837, 6237 ], [ 1200, 3700 ], [ 3600, 5700 ] ] }, { "section_index": 4, "heading": "5 experiment", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14404, 14804 ], [ 0, 1700 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11907.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11896", "assignment_index": 114, "title": "Automated Assembly Instruction Generation from CAD Models Using Grounded Large Language Models: A Human-in-the-Loop Framework", "published": "2026-10-08T13:04:03Z", "authors": [ "Aaron Dsouza", "Mohammed Azeez Khan", "Ashutosh Mishra", "Arshaan Khan", "Neha K. Nair", "Amar Kumar Behera" ], "affiliation": [ "Aaron D’Souza Mohammed Azeez Khan Ashutosh Mishra Arshaan Khan Neha K. Nair Amar Kumar Behera" ], "source_url": "https://arxiv.org/html/2610.11896v1", "source_sha256": "d986b6ab62df897ad20969b2d3e4d087f09811f7742e89c56ed26bf992b51a29", "source_bytes": 249473, "source_json_sha256": "a27b647ed3ed65b3991e0a9d5b6df7fdb8be45082cfe1c3d32ab24e62c6006ca", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "STEP assembly转typedProductGraph，Kahn确定拓扑次序，LLM仅实现selectedgraphslice为自然语言，规则/LLM review标flag；human清完flag才允许PDF。现prototype几何仅bbox/containment，matingconstraints来自提供的assemblystructure。", "evaluation_evidence": "仅内置六part/八constraint referenceassembly单次proof-of-concept，Pillow示意图fallback；展示数量/材料/2.5Nm扭矩和次序保持。没有多assembly、geometrycollision验证、泛化或foundation model方法比较。", "caveats": [ "本篇工程pipeline概念验证，当前foundation-model审查范围外，不宣称已实现CAD能力。", "STEP自动读取kinematic mates、insertion/contact/pose、multimodal generation、simulation均futurework，不能按架构图实线/虚线混同已实现。", "humanreleasegate与groundedslice是重要安全边界，但单reference不能建立正确工程装配指令。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11896v1#S3", "https://arxiv.org/html/2610.11896v1#S4", "https://arxiv.org/html/2610.11896v1#S6", "https://arxiv.org/html/2610.11896v1#S8" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Problem Formulation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2008, 2408 ] ] }, { "section_index": 3, "heading": "4 Proposed Framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11777, 12177 ] ] }, { "section_index": 5, "heading": "6 Proof-of-Concept Case Study", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3782, 4182 ], [ 0, 1600 ] ] }, { "section_index": 7, "heading": "8 Limitations and Open Research Problems", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2935, 3335 ], [ 0, 1400 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11896.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11826", "assignment_index": 115, "title": "From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment", "published": "2026-10-08T12:25:50Z", "authors": [ "Chen Zhao", "Xingping Dong", "Jiachun Shi", "Liang Peng", "Chong Wang", "Zhen Lei", "Ran He", "Bo Du" ], "affiliation": [ "Chen Zhao 1 Xingping Dong 1,† Jiachun Shi 1 Liang Peng 1 Chong Wang 1 Zhen Lei 2 Ran He 2 Bo Du 1 1 School of Computer Science, Wuhan University 2 Institute of Automation, Chinese Academy of Sciences zhaochen.whu2026@gmail.com xingpingdong@whu.edu.cn 2023302111051@whu.edu.cn pengliang@whu.edu.cn cwang@whu.edu.cn zhen.lei@ia.ac.cn rhe@nlpr.ia.ac.cn dubo@whu.edu.cn" ], "source_url": "https://arxiv.org/html/2610.11826v1", "source_sha256": "458bac4700cebeda1f12f9b778d8a352cb6dce39ccb0a82a09106628812db608", "source_bytes": 451298, "source_json_sha256": "d2b913b016c4076dd5ddb28943e5604aad2d1b025bb4b6dc84240bf45d336b80", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ResOT用LURE pairedfaithful/hallucinated完整caption末位hidden作离线calibration；两类SVDtop-r，hallucinatedbasis投影除faithfuldominantsubspace后SVD得residualU。对U内两类均值/收缩cov估计GaussianOT闭式affine；推理logdensityratio sigmoid gate控制alpha和normclip，h+=alpha*g*U(T(U^Th)-U^Th)，正交分量不变。", "evaluation_evidence": "5000LURE训练pairedcaptions，7B LLaVA1.5/mPLUGOwl2/InstructBLIP greedy；POPE、CHAIR/BLEU1、MME、MMHal/AMBER与一般multimodal任务，GPT4o评分COCO描述；不同校准样本量、image-text mismatch/labelswap腐化鲁棒性。", "caveats": [ "训练free仍需离线带faithful/hallucinated标签calibration，不能称完全unsupervised；测试gold不应提供给推理。", "非Gaussian仅保证最小位移affine momentmatching，不保证完整真实分布OT对齐。", "固定完整caption末位校准迁移到逐token解码需实测；faithfulresidual并非0，不能用直接抹除代替OT/gate。", "作者只说codewillbereleased，本轮未查到/未验证作者代码，候选留下一批。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11826v1#S3", "https://arxiv.org/html/2610.11826v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12872, 13272 ], [ 800, 2700 ], [ 2600, 4900 ], [ 5100, 7700 ], [ 7600, 9900 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10859, 11259 ], [ 600, 2100 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11826.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11818", "assignment_index": 116, "title": "From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction", "published": "2026-10-08T12:21:52Z", "authors": [ "Uddipan Basu Bir", "Vincent Christlein", "Andreas Maier", "Mathias Zinnen" ], "affiliation": [ "Uddipan Basu Bir Affiliation: Friedrich-Alexander-Universität Erlangen–Nürnberg, Germany E-mail {uddipan.bir.basu,vincent.christlein,andreas.maier,mathias.zinnen}@fau.de Vincent Christlein Affiliation: Friedrich-Alexander-Universität Erlangen–Nürnberg, Germany E-mail {uddipan.bir.basu,vincent.christlein,andreas.maier,mathias.zinnen}@fau.de Andreas Maier Affiliation: Friedrich-Alexander-Universität Erlangen–Nürnberg, Germany E-mail {uddipan.bir.basu,vincent.christlein,andreas.maier,mathias.zinnen}@fau.de Mathias Zinnen Affiliation: Friedrich-Alexander-Universität Erlangen–Nürnberg, Germany E-mail {uddipan.bir.basu,vincent.christlein,andreas.maier,mathias.zinnen}@fau.de" ], "source_url": "https://arxiv.org/html/2610.11818v1", "source_sha256": "34e7a06e3091ed19d3713d9df6b2af428f5424e61c4937d47b6f1fd2065d6b08", "source_bytes": 330311, "source_json_sha256": "c2371c2e4d156a696b6d0571c2eda7f6b4b912a79eb5c48443efe5a5b5b6146d", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "比较八个200M-7B open VLM真实documentimage→OCR/schemaJSON。按各模型支持接口固定OCRprompt或JSON skeleton，自由prompt模型fewshot仅train示例；<1B全参、较大模型QLoRA。另独立Optuna validationCER调参、经典图像预处理、同epochbudget两阶段learningrate训练与jointcollections。", "evaluation_evidence": "三德国高校collection共1062records，743train/159val/160test；Erlangen308图公开archive、Gemini初稿人工核转写Zenodo。CER/ANLS*无效JSON回落rawtext，mAPF1无效JSON记0；最佳checkpoint最低autoregressive valCER。jointtraining出现数据集间性能转移。", "caveats": [ "固定OCR接口和自由schema接口信息不同，不能把JSON0分简单归因为视觉识字能力弱。", "1062样本单收藏范围，collection/privacy/license与其他两数据是否可公开需下一批逐一核查；仅公开子集可诊断。", "top3选择标准与跨record/page sourcegroup隔离还需查，Optuna5epoch筛选后的fullrun不能用测试再筛。", "institution-controlled/private/sustainable是架构动机，未见生产A/B或直接能耗证据。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/uddipan77/Analysis-of-Lightweight-Vision-Language-Models-for-Document-OCR-and-Structured-Output-Generation", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11818v1#S3", "https://arxiv.org/html/2610.11818v1#S4", "https://arxiv.org/html/2610.11818v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Dataset, Task, and Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10566, 10966 ], [ 700, 2300 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup and Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5669, 6069 ], [ 1100, 3400 ], [ 3400, 5700 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12691, 13091 ], [ 0, 1300 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11818.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11810", "assignment_index": 117, "title": "Seek-and-View Reasoning for Multi-View Spatial Understanding", "published": "2026-10-08T12:18:01Z", "authors": [ "Qixiang Chen", "Cheng Zhang", "Fucai Ke", "Chi-Wing Fu", "Jianfei Cai", "Jingwen Ye" ], "affiliation": [ "Qixiang Chen Cheng Zhang Fucai Ke Chi-Wing Fu Jianfei Cai Jingwen Ye Affiliation: Monash University Affiliation: The Chinese University of Hong Kong" ], "source_url": "https://arxiv.org/html/2610.11810v1", "source_sha256": "8e3047e24aa359a0014b855286b8ab6bed02789c0934a59f9aa9f7fc37bf133c", "source_bytes": 634224, "source_json_sha256": "1af26f3eb9158a240b38c846c29799e82727f56db617702405dfe0a28f2287f0", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Vantage trainingfree VLM根据问题/原视图选reference、解释directionframe并从17相机action×3magnitude规划；G3T gravity-aligned3D reconstruction/pose将语义action映射metricpose并reproject合成question-relevantview。最终VQA同时保留原图、新图、analysis和guide，不用模型假想图替代真实3D几何。", "evaluation_evidence": "六VLM、5空间bench子集：MindCubeTiny、MMSI position/attribute、BLINK multiview，以及OmniSpatial perspective/SPIN rotationtranslation；H200×2 G3T与openVLM，accuracy与componentablation。", "caveats": [ "合成视图可能不完整/几何重建失败，最终VLM仍可能忽略证据，作者做失败分析。", "SPIN中物体动态但背景静态；不能推广到一般时空动态场景。", "extraVLMplanning+3DFM两GPU有成本，须与同compute/textcontext对照；6backbone无fine-tuning不代表零额外资源。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/q1xiangchen/Vantage", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11810v1#S3", "https://arxiv.org/html/2610.11810v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10792, 11192 ], [ 1400, 4100 ], [ 5200, 8500 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13099, 13499 ], [ 650, 1800 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11810.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11776", "assignment_index": 118, "title": "DPPM: Dual-Path Parametric Memory for Personalized Language Models", "published": "2026-10-08T11:56:07Z", "authors": [ "Yuhao Chen", "Shuochen Liu", "Jiayao Shi", "Jian Hong", "Chen Cheng", "Xinyun Ding", "Tao Wang", "Ya Li", "Quan Liu", "Tong Xu" ], "affiliation": [ "Yuhao Chen Affiliation: University of Science and Technology of China Email: isyuhaochen@mail.ustc.edu.cn Shuochen Liu Affiliation: University of Science and Technology of China Jiayao Shi Affiliation: University of Science and Technology of China Jian Hong Affiliation: University of Science and Technology of China Affiliation: iFLYTEK Research Group Chen Cheng Affiliation: iFLYTEK Research Group Xinyun Ding Affiliation: iFLYTEK Research Group Tao Wang Affiliation: iFLYTEK Research Group Ya Li Affiliation: iFLYTEK Research Group Quan Liu Affiliation: iFLYTEK Research Group Tong Xu Affiliation: University of Science and Technology of China" ], "source_url": "https://arxiv.org/html/2610.11776v1", "source_sha256": "8ec004c5cd91a52e2559bc46ae9af6a32bc9a59bd777bbfe0e52336805077c89", "source_bytes": 313472, "source_json_sha256": "adb44fcb258890842bbb7cab90fb4d9902716f70370e66e3935e7600790fd287", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "DPPM由frozen compiler编码historysegment且排除query/answer；Evidence coordinatewise temporalsoftmax直接池化并RMS校准，Delta用normalizedELUkey、sigmoidgate对associativestate写预测残差并平行onesnormalizer。learnedfuse两路径，经frozen decoder生成historyconditionedLoRA，与sharedadapter相加；answerCE仅训练memorymodules。", "evaluation_evidence": "PersonaMemv2 18527train2059val5000test且testpersona不入train/val；PrefEval seed42 16train/4heldouttopics，7380train1620test、10/70/300turn，2B-7Bbackbones；报告accuracy及Evidence/Delta/目标ablation，额外FKL/GRPO并非全部改善。", "caveats": [ "直接池化下界仍随historylength下降，只缓解ageattenuation不是消除遗忘。", "associativekey相似度会干扰，Evidence本身orderinvariant，须真实Delta/Fusion/RMS/decoderLoRA而非摘要prompt代理。", "训练监督可读trainanswer，推理memory必须排除当前query/answer；个性化隐私、跨session污染与多模态扩展未验证。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/isyuhaochen/DPPM", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11776v1#S3", "https://arxiv.org/html/2610.11776v1#S4", "https://arxiv.org/html/2610.11776v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7301, 7701 ], [ 700, 4100 ], [ 4100, 6500 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13059, 13459 ] ] }, { "section_index": 6, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 384 ], [ 0, 384 ], [ 0, 384 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11776.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11766", "assignment_index": 119, "title": "Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation", "published": "2026-10-08T11:50:32Z", "authors": [ "Haitong Jiang", "Chunlin Liu", "Sihan Tang", "Chan Wu", "Xiaoqing Su", "Yuhong Feng" ], "affiliation": [ "Haitong Jiang Affiliation: College of Computer Science and Software Engineering, Shenzhen University Chunlin Liu Affiliation: College of Computer Science and Software Engineering, Shenzhen University Sihan Tang Affiliation: School of Intelligence Science and EngineeringHarbin Institute of Technology (Shenzhen), Shenzhen, China * Corresponding author: yuhongf@szu.edu.cn Chan Wu   Xiaoqing Su   Yuhong Feng Affiliation: College of Computer Science and Software Engineering, Shenzhen University" ], "source_url": "https://arxiv.org/html/2610.11766v1", "source_sha256": "b8f766d9c46819eeb94c9e9b93334f7a2af915c1faa4a82b84439ed38a7c2363", "source_bytes": 221247, "source_json_sha256": "9b28fbbc6d6bce1b31766eda8d95448c23550994017d32edece9074cfb79db04", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "IRIS将operative understandingrate UR与ASR并报；UR要求识别且把目标作为正在回答的任务，task-aware refusal/safealternative算recovered，纯翻译/语言分析不算。分开taskrecognition、responseframe、harmfulassistance、blocked/empty，后两仍计两denominator。", "evaluation_evidence": "105去重可跨接口有害任务、六closed/API模型，English direct与semanticcompressed接口及渐明确English reconstruction；similarASR对应不同UR，FormalLogic高UR也可高ASR。人类agreement/judge敏感度与bootstrap量化标签不确定性。", "caveats": [ "UR是response-based reader-dependent指标，不是内部理解或新的safetyscore；bootstrap条件于标签。", "105精选panel不代表真实攻击分布，T1-T3显式化改变多因素，不能因果归因只有recovery。", "下一批只实现授权安全评测统计/标注协议；不直接执行或发布危险任务答案，provider版本漂移需固定metadata。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/kevinjiang0121-cyber/IRIS", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11766v1#S2", "https://arxiv.org/html/2610.11766v1#S3", "https://arxiv.org/html/2610.11766v1#S4", "https://arxiv.org/html/2610.11766v1#Sx1" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Recovery-Conditioned Safety Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1153, 1553 ], [ 0, 1553 ] ] }, { "section_index": 2, "heading": "3 Empirical Evidence", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5305, 5705 ] ] }, { "section_index": 3, "heading": "4 Discussion and Conclusion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1038, 1438 ] ] }, { "section_index": 4, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 365, 765 ], [ 0, 765 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11766.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11761", "assignment_index": 120, "title": "RAGenome: Scaling Retrieval-Based Genomic Language Models to Long Contexts", "published": "2026-10-08T11:44:58Z", "authors": [ "Frederikke Isa Marin", "Panagiotis Antoniadis", "Dionysia Danai Brilli", "Andreas Bjerregaard", "Rachael DeVries", "Yan Li", "Ole Winther", "Wouter Boomsma" ], "affiliation": [ "Frederikke Isa Marin   Panagiotis Antoniadis 1 1 footnotemark: 1 Dionysia Danai Brilli † † thanks: Equal contribution. Affiliation: University of Copenhagen Andreas Bjerregaard   Rachael DeVries   Yan Li   Ole Winther 2 2 footnotemark: 2 Affiliation: University of Copenhagen Affiliation: Novo Nordisk A/S Affiliation: Technical University of Denmark Wouter Boomsma † † thanks: Corresponding authors: ole.winther@bio.ku.dk, wb@di.ku.dk Affiliation: University of Copenhagen" ], "source_url": "https://arxiv.org/html/2610.11761v1", "source_sha256": "18f058ca100e3fe72cc7f1c5256022c4f71d9f0388988d75cfa9cf5b0475334f", "source_bytes": 176294, "source_json_sha256": "6d58f2f32af694e46aff20839025f3d16190967ace4e07e315d13d8610f6c5ab", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "RAGenome读取100vertebrate whole-genomealignment，丢gap却保留originalalignmentcolumn RoPE，按phylogenetic tree固定tokenbudget挑lineage，taxonomy各级共享embedding。独立species slidingwindowlocalattention+middlequery对homologcrossattention，clade masking MLM、分阶段扩展13312context、168M模型。", "evaluation_evidence": "BEND genefindingfrozenembedding+two-layerCNN、官方test MCC约.45→.60；ClinVar zero-shot variantprioritization维持演化任务能力；最大retrievalbudget80000适配H100，较MSA方法减少gap训练token。", "caveats": [ "推理必须有WGA，不能作用无alignment序列/物种；只humanquery训练、单168M规模。", "跨query/homolog位置与clademasking都是核心，不能用普通检索text或无alignment单seqLM代理。", "这是公开基因组representation诊断候选，不提供临床诊断或sequence设计；成本比较须含alignment/retrieval获取。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/PanosAntoniadis/RAGenome", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/pantoniadis/RAGenome", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11761v1#S3", "https://arxiv.org/html/2610.11761v1#S4", "https://arxiv.org/html/2610.11761v1#S5", "https://arxiv.org/html/2610.11761v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Related Work", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3178, 3578 ] ] }, { "section_index": 3, "heading": "4 RAGenome", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10047, 10447 ], [ 650, 3700 ], [ 3800, 6800 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 9351, 9751 ], [ 0, 1800 ] ] }, { "section_index": 5, "heading": "6 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 656 ], [ 256, 656 ], [ 0, 656 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11761.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11674", "assignment_index": 121, "title": "VESSI - VLM-Enhanced Support for Surveillance and Investigations", "published": "2026-10-08T10:50:17Z", "authors": [ "Saverio Cavasin", "Pietro Tedeschi", "Mattia Tamiazzo", "Alessandro Brighente", "Simone Milani", "Mauro Conti" ], "affiliation": [ "Saverio Cavasin Pietro Tedeschi Mattia Tamiazzo Alessandro Brighente Simone Milani Mauro Conti † † thanks: S. Cavasin, A. Brighente, and M. Conti are with the Department of Mathematics, University of Padua, Padua, Italy. † † thanks: S. Cavasin and P. Tedeschi are with the Innovation & Research Center, CY4GATE S.p.A., part of ELT Group, Rome, Italy. † † thanks: M. Tamiazzo and S. Milani are with the Department of Information Engineering, University of Padua, Padua, Italy. † † thanks: M. Conti is also with Örebro University, Örebro, Sweden. † † thanks: Corresponding author: Saverio Cavasin (e-mail: saverio.cavasin@phd.unipd.it). E-mail addresses: pietro.tedeschi@cy4gate.com (P. Tedeschi), alessandro.brighente@unipd.it (A. Brighente), simone.milani@dei.unipd.it (S. Milani), and mauro.conti@unipd.it (M. Conti)." ], "source_url": "https://arxiv.org/html/2610.11674v1", "source_sha256": "cdd522f9e350ccf9239e8e4d713bba393d38e8545c27ff94b29c57ff2ca0040f", "source_bytes": 596762, "source_json_sha256": "062438fa1b0947c344fddefe1319dfdfa5232941366d2c2a1bb26077ce5bee1a", "status": "rejected_or_out_of_scope", "qualifies_for_reproduction_consideration": false, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": false, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "VESSI按已知video类别给独立抽帧VLM category-specificprompt，response归Yes/No/Unclear，平均生成token最大softmax只作within-modelsalience筛选。CMUS组合coverage、相对视频时长processingtime、Yesrate及固定Yesratebucket惩罚，权重1/1/-.5/-1。", "evaluation_evidence": "UCF-Crime950视频/13类、四约7Boff-the-shelf模型，manual65videos5909frames与50normalvideo补充排序检查；原文报告>66%video至少flag、>85%时间节约，但不是完整frameGT或真实人工reviewtime试验。", "caveats": [ "主要为执法监控应用pipeline/operationalproxy，不在当前foundation-model方法范围，不新增实现。", "knownvideocategoryprompt和independentframes缺乏temporalcontinuity；Yesrate惩罚不证明hallucination，softmax不是crimeprobability。", "不推广为个人犯罪倾向/司法判断或无人审核部署；原文也要求humaninterpretation，当前review只分析公开论文。", "timeSaved是processingvsoriginalduration，不等实际人类审阅时间或生产A/B。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/SvrCvs/vlm_surv", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/huggingface/transformers/releases/tag/v4.49.0", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/huggingface/transformers/releases/tag/v4.49.0.dev0", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/pytorch/pytorch/releases/tag/v2.1.2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/pytorch/pytorch/releases/tag/v2.6.0", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11674v1#S3", "https://arxiv.org/html/2610.11674v1#S4", "https://arxiv.org/html/2610.11674v1#S5", "https://arxiv.org/html/2610.11674v1#S6", "https://arxiv.org/html/2610.11674v1#S8" ], "evidence_sections": [ { "section_index": 2, "heading": "III Tested VLM", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3975, 4375 ] ] }, { "section_index": 3, "heading": "IV The VESSI Framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7914, 8314 ], [ 650, 2000 ] ] }, { "section_index": 4, "heading": "V Experimental Protocol", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1590, 1990 ], [ 0, 1400 ] ] }, { "section_index": 5, "heading": "VI Composite Model Utility Score ( CMUS )", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5999, 6399 ], [ 0, 1800 ], [ 3400, 6399 ] ] }, { "section_index": 7, "heading": "VIII Limitations and Future Work", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1919, 2319 ], [ 0, 1200 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11674.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11610", "assignment_index": 122, "title": "Beyond Report Imitation: Clinically Aware Multi-Image Ultrasound Report Generation from Visible Evidence", "published": "2026-10-08T09:53:46Z", "authors": [ "Yuchen Yang", "Xin Wang", "Lufan Wang", "Yinghong Pan", "Yujuan Feng", "Yuqing Yang" ], "affiliation": [ "Yuchen Yang Wang Xin Lufan Wang Yinghong Pan Yujuan Feng Yuqing Yang" ], "source_url": "https://arxiv.org/html/2610.11610v1", "source_sha256": "52d6342c78a32fe56783c96f832f00220732593c659c5a9dce95d49a94cfa459", "source_bytes": 226770, "source_json_sha256": "18b0963d4a6cf844d268786595de7cf953366cd7ee810a1638fbcd8666670691", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "CAMEO将两张archivedultrasound原报告拆atomicpoint，以两个LLMverifier+arbiter分visible/partial/unsupported/excluded和supportview构造trustedtemplate，visible正例、partial qualified。StageIprojector对齐+singleimageadapt、StageIIcrossviewQA/reportSFT、StageIII临床错误chosen/rejected的DPO；visionencoder冻结，后两阶段更新projector/LLM。", "evaluation_evidence": "PubMedVision-US instruction25631、USReportDistilled5890train/1474test及17670traininginstances、Pref21869train1354val不作为testreference。ClinicalScore55.02→74.20；blindreader301atomicannotationsteacheraccuracy78.9%/F184.8%，rawpoint25.2%unsupported。StageII主要lexicalgain、StageIII重塑errorprofile。", "caveats": [ "两LLM一致+arbiter是弱监督净化，不是放射科标注；同template用于train/evaluation存在teacherdefinition耦合，需独立reader审计。", "两视图三器官不涵盖完整动态scan/rareconditions，ClinicalScore仍proxy，不宣称clinicaldeployment。", "nextbatch仅publicdeidentifiedsource评估/机制，不生成患者实际诊断；case级split与teacher/test信息隔离须核。", "不能以空refusal/正常模板低hallucination当可靠性提升，应并报omission/misdiagnosis/overcompletion。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/NiHaoWoJiaoYYC/CAMEO", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11610v1#S3", "https://arxiv.org/html/2610.11610v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 2000, 4900 ], [ 4900, 9100 ], [ 9000, 12000 ], [ 11900, 14000 ], [ 14000, 15200 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15262, 15662 ], [ 650, 1800 ], [ 11900, 14400 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11610.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11605", "assignment_index": 123, "title": "NanoProof: Open and Efficient Automated Theorem Proving in Lean 4", "published": "2026-10-08T09:49:39Z", "authors": [ "Matěj Kripner", "Milan Straka" ], "affiliation": [ "Matěj Kripner 1 1 footnotemark: 1 Milan Straka Affiliation: Institute of Formal and Applied Linguistics, Faculty of Mathematics and Physics, Charles University , Prague , Czech Republic" ], "source_url": "https://arxiv.org/html/2610.11605v1", "source_sha256": "454a500d095ec3205b70847267b9f4f111faa44aa2aa5554be08826e70c65bab", "source_bytes": 304361, "source_json_sha256": "aadd07c5f79f9a170dede542d03d122e44c9a0fc2b3e8d67053d542756dfb283", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "NanoProof factorizedAND-OR MCTS actor生成Lean tactic、critic预测negativeproofsubtreedepth，Lean立即执行丢invalidtactic，所有分支需闭合；matchmaker重加权能力边缘theorem，successfulproofreplay混SFT循环训练。LeanTree从人写proof拆结构tree、支持中间state重启与simplification/diversetacticsampling，最终导出standaloneLean独立验证。", "evaluation_evidence": "10B Nemotronmathbasetrain、Mathlib287kproofsteps文件最后10%29kval、258ktrain；MiniF2FValid手调参，Testfinal，三independentrunsmin/max。报告Test50.8%pass@16及约90x/7xclassbaselinecompute下降，hardware/FLOP部分由作者估算。", "caveats": [ "更强openweightprover存在但不是相同factorizedfromscratch类别；不可claim全部theoremproverSOTA。", "LeanREPL曾接受invalidproof，作者记录修正，必须standaloneverifier且不允许sorry/admit。", "训练proof/criticdepth监督合法，评估agent只接theorem/currentgoals，不读goldproofplan。", "compute估算非统一实测、pass@16/searchsimulations定义需下一批完整协议核验。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Paper-Proof/paperproof", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/facebookresearch/Evariste", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/karpathy/nanochat", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/kim-em/lean-training-data", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/kripner/leantree", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/kripner/nanoproof", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/leanprover-community/repl", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/leanprover-community/repl/blob/master/test/have_by_sorry.in", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/AI-MO/NuminaMath-LEAN", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/ufal/leantree", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/openai-community/gpt2", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11605v1#S3", "https://arxiv.org/html/2610.11605v1#S4", "https://arxiv.org/html/2610.11605v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3. NanoProof Automated Theorem Prover", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10253, 10653 ], [ 700, 3100 ], [ 4400, 7100 ] ] }, { "section_index": 3, "heading": "4. Structured Proof Tree Dataset", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6857, 7257 ] ] }, { "section_index": 4, "heading": "5. Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3454, 3854 ], [ 0, 1400 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11605.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11592", "assignment_index": 124, "title": "Probing for Long-Horizon Deductive Reasoning Capabilities in Language Models with Prolog", "published": "2026-10-08T09:42:07Z", "authors": [ "Hadeel Al-Negheimish", "Jasna Ilieva", "Yoon Kim" ], "affiliation": [ "Hadeel Al-Negheimish Affiliation: King Saud University Affiliation: Massachusetts Institute of Technology Email: hadeelan@mit.edu Jasna Ilieva Affiliation: Massachusetts Institute of Technology Email: jasna23@mit.edu Yoon Kim Affiliation: Massachusetts Institute of Technology Email: yoonkim@mit.edu" ], "source_url": "https://arxiv.org/html/2610.11592v1", "source_sha256": "5833142aea2a4304c6b87492bb821de84bb7c28673671b8ef7f2bc19335d789f", "source_bytes": 666191, "source_json_sha256": "53b4dcb1b7eae81534e6e1257b295c6af4ab9d4a3dd48f41e5557a623711b77e", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ProloNg合成无loop/recursion/negation、arity≤2、eachpredicateone rule的definitePrologDAG；GenerateRule保证一条深Npath、Poissonbodybreadth，生成pairedprovable/unprovablequeries并mutation单fact、shuffle全部rule/fact。sharedcontext扩展同program不同depthquery隔离长度。", "evaluation_evidence": "8reasoningmodels、depth至22/62kcontext，多数depth>10近balancedchance.5；原depth/lengthr.97 GLMM无法可靠拆因果。另50sharedcontexts×6depth×正负共600samples，r.0003，四openmodels两因素均显著影响。", "caveats": [ "这是最简单无backtracking单链/AND树演绎，不代表一般Prolog或真实复杂逻辑推理。", "闭源reasoningsummary不是完整trace；64k/82k输出budget仍可能影响深层失败。", "sourceprogram/query供model，label仅verifier读取，不能把GenerateRule goldderivation混进prompt；negative mutation需独立Prolog判真。", "原correlatedGLMM结论与sharedcontext证据须区分，不按abstract把depth独立归因。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/halnegheimish/ProloNg", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11592v1#S3", "https://arxiv.org/html/2610.11592v1#S4", "https://arxiv.org/html/2610.11592v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 ProloNg : Prolog Long Reasoning", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5851, 6251 ], [ 1200, 4000 ] ] }, { "section_index": 3, "heading": "4 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 7204, 7604 ], [ 4600, 6900 ] ] }, { "section_index": 5, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 323, 723 ], [ 0, 723 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11592.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11585", "assignment_index": 125, "title": "Adapting English Quality Classifiers for Multilingual LLM Pretraining Data Selection", "published": "2026-10-08T09:33:32Z", "authors": [ "Vinko Sabolčec", "Bettina Messmer", "Yassine Turki", "Martin Jaggi" ], "affiliation": [ "Vinko Sabolčec Bettina Messmer Yassine Turki & Martin Jaggi Affiliation: EPFL Email: firstname.lastname@epfl.ch" ], "source_url": "https://arxiv.org/html/2610.11585v1", "source_sha256": "1895e6c55e67ff63a8dad5b9e31c8bad2c984f13ae78f81ff77214eb721160a2", "source_bytes": 268693, "source_json_sha256": "64aa3dbcd500250f67cca1dc6d4153ce5c12b985e59408553f14b5d3586ea74f", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "将FineWeb2 top100language-scriptpair原文0.1%样本Qwen3-32B译English，再EnglishFineWebedu/DCLM教师打分；frozenXLMR/mmBERT原语言embedding上two3072ReLUhiddenMLP回归，edu[0,5]L1、DCLMclipprob后logitL1。约6M样本包含1MEnglish/5Mmulti，top10%perlanguage过滤并保持原languagepretrainmix。", "evaluation_evidence": "Apertus架构1B/3B/8B、4096context、AdEMAMix/WSD、约33kH100hourspretrain；译文约750H100hours。对比multilingualfilterbaseline、regional/culturalknowledge及未训练语言syntheticvswebteacher相关，报告相近downstream表现。", "caveats": [ "教师English质量criteria通过翻译继承，不是语言/文化中立groundtruth；machine translation错误/风格可能系统改变标签。", "topN按language-scriptbytecount不是语言人数；originallanguageembeddinggeneralization不保证所有1800pair已真实训练验证。", "下一批需冻结原doc/translation/teachers、languagegroupheldout、仅validation阈值；33kGPUhour完整pretrain超本批范围，不以少量classifierloss充当LMqualitygain。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/huggingface/datatrove", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-32B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3-8B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/epfml/FineWeb2-HQ-PlusPlus-Classifier", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/google/gemma-3-27b-it", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/mlfoundations/fasttext-oh-eli5", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/openai/gpt-oss-120b", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/spaces/HuggingFaceFW/blogpost-fine-tasks", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11585v1#S3", "https://arxiv.org/html/2610.11585v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4057, 4457 ], [ 650, 3100 ], [ 3000, 4457 ] ] }, { "section_index": 3, "heading": "4 Results and Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 650, 1800 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11585.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11548", "assignment_index": 126, "title": "Conditional Transfer from Controlled Pretraining Mixtures to Code", "published": "2026-10-08T09:12:45Z", "authors": [ "Ohad Rubin" ], "affiliation": [ "未核实；官方HTML作者单位块缺失，缓存提取的作者姓名不是机构" ], "source_url": "https://arxiv.org/html/2610.11548v1", "source_sha256": "d429903301ae1d95cc2673c9e79cd84f75618cce328db43b169447b4e813f44d", "source_bytes": 230450, "source_json_sha256": "80b0d15fcb18c74bc60a53c537ec2e5866452ee4fd136327ee4b6deaa1d72750", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "固定70%CodeParrotPython+30%controlledsource simplex(OpenCodeInstruct/12curatedsynthetic/15literatureprobe)，固定模型/训练/Ft桥，分别定义diagnosticloss跟globalprogress、teachableloss跟ownanswer-tokenbudget、source对HumanEvaltransfer。Ado按近期loss拟合scalinglaw估未来1000step可下降量调source/taskprob。", "evaluation_evidence": "15coarsemixtures、60k/120kstepssparseexpertLM及固定1000updatecodeSFT；14/27tasksteachable但ranking不同于transfer。最优observedcontrolledslice.75OpenCode+.25curated pass20 22.6，纯Open19.5；三free-ratioAdo末Open仅1.2-1.4%、劣fixedcontrol2.4-11pp。", "caveats": [ "主endpoint对HumanEval checkpoints100..1000取best，作者承认optimistic.2-1.6pp；bootstrap重做selection不将其变为独立test比较，下一批验证选checkpoint后isolatedtest。", "OpenCodepretrain与FT共享distribution，format/content同时改变；teachability在mixturecoupledsweep仅association不是taskisolatedcausal。", "15pointgrid只能bestobserved非真实argmax，单architecture/target/scheduler不可普遍否定adaptivecurriculum。", "原HTML作者block缺明确单位，affiliation不按作者姓名当机构。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11548v1#S3", "https://arxiv.org/html/2610.11548v1#S4", "https://arxiv.org/html/2610.11548v1#S7", "https://arxiv.org/html/2610.11548v1#S9" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4462, 4862 ], [ 700, 4000 ] ] }, { "section_index": 3, "heading": "4 Three Signals", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3894, 4294 ] ] }, { "section_index": 6, "heading": "7 Adaptive Scheduling Optimizes Reducibility and Loses Transfer", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4358, 4758 ], [ 0, 1600 ] ] }, { "section_index": 8, "heading": "9 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1152, 1552 ], [ 0, 1400 ], [ 1300, 1552 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11548.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11510", "assignment_index": 127, "title": "Does Modern Standard Arabic (MSA) Dominate Arabic Dialects in LLMs? A Representation-Level Analysis", "published": "2026-10-08T08:46:47Z", "authors": [ "Abdu Sallouh", "Nicholas Popovič", "Michael Färber" ], "affiliation": [ "Abdu Sallouh Affiliation: Technical University of Dresden Email: abdu.sallouh@mailbox.tu-dresden.de Nicholas Popovič † † thanks: Work conducted while at TU Dresden & ScaDS.AI Affiliation: Mohamed bin Zayed University of Artificial Intelligence Email: nicholas.popovic@mbzuai.ac.ae Michael Färber Affiliation: Technical University of Dresden Affiliation: ScaDS.AI, TU Dresden Email: michael.faerber@tu-dresden.de" ], "source_url": "https://arxiv.org/html/2610.11510v1", "source_sha256": "aa20200cb7b547b1b09271393e2af0ae2822bd3a44f81051be64763bbbf11ccd", "source_bytes": 103574, "source_json_sha256": "982e81faded8305425292c18c299698dd80fed04c1b0e705166afa8c13d155ba", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "在MADAR26Arabicvarieties语义parallel句抽各层hidden，NMI测dialectlabel/clusterseparability；用true-dialect与每个competitor alignedtokenposterior之比Lambda，汇总所有token/competitorpair中Lambda<1比例，不聚合单token竞争方向、不把低NMI等价MSAdominance。", "evaluation_evidence": "BLOOM/mGPT每variety400句、附加model100句，Arabic平均NMI约.1/.2但confusion仍diagonal、linearprobe可预测；noevidenceMSAdominance，architecture separabilitypeak可后移，POS与首tokenablation。", "caveats": [ "travel域、小writer数量可能编码writer而非dialect；同positionalignedtoken不自动保证跨tokenizer语义对齐，需下一批核对token匹配实现。", "无dominance证据不等从不存在、也不否定输出MSAbias；geometrydiagnostic不改真实generation。", "rule-basedPOStagger不完善，400/100samples探索性结论不能推广全部Arabic模型。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/abduABS/MSA_Dominance_LLM", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11510v1#S3", "https://arxiv.org/html/2610.11510v1#S4", "https://arxiv.org/html/2610.11510v1#S6" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Experimental Framework", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4890, 5290 ], [ 650, 3500 ] ] }, { "section_index": 3, "heading": "4 Results and Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4972, 5372 ] ] }, { "section_index": 5, "heading": "6 Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1479, 1879 ], [ 0, 1300 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11510.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11508", "assignment_index": 128, "title": "WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models", "published": "2026-10-08T08:44:26Z", "authors": [ "Junmyeong Lee", "Dongmin Shin", "Min-Gyu Park", "Wooseok Jeon", "Inho Chang", "Hae-Gon Jeon" ], "affiliation": [ "Junmyeong Lee Affiliation: Yonsei University. Dongmin Shin Affiliation: Yonsei University. Affiliation: Korea Electronics Technology Institute(KETI). Min-Gyu Park Affiliation: Korea Electronics Technology Institute(KETI). Wooseok Jeon Affiliation: Yonsei University. Inho Chang Affiliation: Korea Electronics Technology Institute(KETI). Hae-Gon Jeon † † thanks: ∗ Equal Contribution. † † thanks: † Corresponding Author. Affiliation: Yonsei University." ], "source_url": "https://arxiv.org/html/2610.11508v1", "source_sha256": "3b79d62544015202c5b690f973cd0af18985e006ae8e3b24c19a4f2e89b96ef7", "source_bytes": 195719, "source_json_sha256": "1bef83fa33626c094ab09855011589018475acc7ba7a6ebecbc8ccb5a93ec375", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "WARP-VLA冻结visionencoder/actionpolicy，在wristtoken接口MoE恢复canonicalfeatures；训练canonicalmeananchor作KV、perturbedtokens作Q的sharedcrossattentiontransformer加15expertLoRA。router用residual/anchormean,std,x/yspatialmoments，13translation+2tilt组内softmax各占1/2；每expertfeatureMSE及基于routing×exp(-error)×poseprior的stopgradresponsibility监督router。", "evaluation_evidence": "2000LIBEROdemonstrations40tasks重render10perturbations，训练8RTXPRO6000约30h；LIBERO/pi.5平均39.2→78.3%，LIBEROPlus也改善。实xArm6十task/每mount5trials，basepolicy在500realcanonicaldemo微调，adapter固定simulationtrained但重新计算realcanonicalanchor。", "caveats": [ "knownposeprior/canonicaltargets只训练，推理不得读真cameraextrinsics；不是unsupervisedadapter。", "sim-to-real不是完全zero-data系统，真实basepolicy与anchor需canonicalreal demonstrations，不省该资源。", "下一批只simulationpublicdiagnostic，真实机器人需独立授权、安全场地；少量mount/trials不能称生产A/B。", "不能以imagewarping或templateaction替代nativeVLAfeatureMoE/crossattention真实policy执行。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11508v1#S3", "https://arxiv.org/html/2610.11508v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "III THE PROPOSED METHOD", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 13810, 14210 ], [ 1000, 3900 ], [ 3800, 8200 ], [ 8200, 11500 ], [ 11400, 13100 ] ] }, { "section_index": 3, "heading": "IV Experimental results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 12071, 12471 ], [ 650, 2000 ], [ 6200, 8500 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11508.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11469", "assignment_index": 129, "title": "SAGE: Sink-Aware Guided Emphasis for Visual Grounding in Vision-Language Decoders", "published": "2026-10-08T08:18:58Z", "authors": [ "Jeonghyo Song", "YoungJoon Yoo" ], "affiliation": [ "Jeonghyo Song * Email: thd9592s@cau.ac.kr YoungJoon Yoo Department of Artificial Intelligence, Chung-Ang University Email: yjyoo3312@cau.ac.kr * First author. Corresponding author" ], "source_url": "https://arxiv.org/html/2610.11469v1", "source_sha256": "bf4d760560bd39616145af718dc41d0f43c978074cd233803c5406784a79218b", "source_bytes": 380280, "source_json_sha256": "3aeae4e262ab125ac1c2a6b09e30e6df6f89112eccc3d3f592b6788a87eb6605", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SAGE对nativeattentionpre-softmax视觉keylogits在tokenalignedROI内+alpha5、ROI外-beta2；CLIPpatch-textcosine可queryconditioned，ViTrollout/DINOv3CLSattention是queryagnostic saliency，topk nearest-neighbor映射visualgrid。singlelayersweep诊断PIS与spectralconcentration，可做topklayers。", "evaluation_evidence": "LLaVA1.5/Qwen2VL/InternVL3/NVILA多bench；主表每backbone/bench/ROIsource取最佳sweeplayer是upperbound。另LLaVA7B先按label-freeSinkIoU/top1spectralminimum选固定L24，再所有bench统一评估；TextVQA/GQA收益较大，其他小，same-size randomROI无收益。", "caveats": [ "主表besttestlayer不可当部署validation-selected结果，必须与fixedL24protocol区分。", "只有CLIPmask依query，不能把全部ROIsource叫querydependent；碎/错ROI可能抑真实证据。", "nextbatch需真实nativeattention/tokenalignment及独立validation选层、originaldata，不以goldbbox/oracleROI做能力实现。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11469v1#S3", "https://arxiv.org/html/2610.11469v1#S4", "https://arxiv.org/html/2610.11469v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 14211, 14611 ], [ 1700, 4600 ], [ 5200, 8300 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 10096, 10496 ], [ 600, 2100 ], [ 8400, 10200 ] ] }, { "section_index": 5, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2290, 2690 ], [ 0, 1300 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11469.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11454", "assignment_index": 130, "title": "Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains", "published": "2026-10-08T08:09:17Z", "authors": [ "Miruna Cretu", "Alex Abrudan", "Antonia Panescu", "Tynan Perez", "Rishabh Anand", "N. Benjamin Erichson", "Michael W. Mahoney", "Samuel Blau", "Joseph Jacobson", "Rafael Gómez-Bombarelli", "Rex Ying", "Tuomas Knowles", "Pietro Liò", "Alex Morehead" ], "affiliation": [ "未核实；官方HTML作者单位块缺失，缓存提取的作者姓名不是机构" ], "source_url": "https://arxiv.org/html/2610.11454v1", "source_sha256": "e8bc1a9415480e12cceb250cbed5c9235776730619c2f65a6b84cee02285b4b2", "source_bytes": 728251, "source_json_sha256": "1acf63d79122f954fbaa02846c65741514cd199b84f6777ebbf46ba7fea6e853", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "Zatom2 atom-token-atom U型localencoder/globaltokentrunk/localdecoder，molecule/materialatom1与proteinresidueatom14共享conditionalflowmatchinggeometry；time-gatedelement/sequence/fractional/latticeheads、optionalenergy/forceheads，多任务与跨OMol/OMat预训练，globalmeanforcenorm条件控制relaxationregime。", "evaluation_evidence": "OMol25 3,986,754+OMat24 1,009,850约5M structures，80epochs多配置；UMA-S1p2计算生成样本forceconsistencyAUROC/MAE、分布fidelity，公开molecule/materialbench；2000protein-domainfinetune lengthextrapolationdesignability67.8→74.8%。", "caveats": [ "directforcehead既不强制energy-gradientconsistency也非exactrotationequivariance，不能当物理约束保证。", "UMA评估是learnedpotentialproxy非新DFTgroundtruth；需检查pretrainedpotential与train/testoverlap及geometryvalidity。", "下一批限定通用公开无害atomisticrepresentation/materialdiagnostic；不开展有害分子、毒素或病原体设计。", "当前HTML作者单位block缺失，机构归属未核实，不按AlexMorehead姓名当affiliation；5M/80epochs完整预训练不在本16篇范围。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/RosettaCommons/foundry/tree/b02eed6a6bdf8f44d14a80cc36e3da13c9f2291c/models/rfd3", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/TyJPerez/AtomisticGenEvals", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/Zatom-AI/nucleus", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Ty-Perez/ct-scd-amp20", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/Ty-Perez/ct-scd-omol25", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/LeMaterial/LeMat-Bulk", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/chaitjo/MP20_ADiT", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/facebook/OMAT24", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/facebook/OMol25", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11454v1#S2", "https://arxiv.org/html/2610.11454v1#S3", "https://arxiv.org/html/2610.11454v1#S4" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Preliminaries", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 6348, 6748 ], [ 0, 1300 ], [ 1300, 3000 ] ] }, { "section_index": 2, "heading": "3 Zatom-2", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5608, 6008 ], [ 700, 4200 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 650, 2600 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11454.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11430", "assignment_index": 131, "title": "BioBigBird: A Sparse Attention Model for Long-Range Dependency Processing in Biomedical Text", "published": "2026-10-08T07:54:12Z", "authors": [ "Roshan Balaji", "Pavan Kumar S", "Vasudev Gupta", "Sreejith N", "Keerthana Sridhar", "Nirav Bhatt" ], "affiliation": [ "Roshan Balaji Pavan Kumar S Vasudev Gupta Sreejith N Keerthana Sridhar Nirav Bhatt Affiliation: Wadhwani School of Data Science and AI, Affiliation: Indian Institute of Technology Madras, Chennai, India" ], "source_url": "https://arxiv.org/html/2610.11430v1", "source_sha256": "49636dbd8c15d9b55e6548996c05590df4ed7049362bc743e011ccbe28af7b36", "source_bytes": 148422, "source_json_sha256": "280988c21e4652f7670d0f533ab87a43809a8aa606b8b5014f4cc08b3ef5e67b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "BioBigBird沿用BigBird稀疏双向4096context，PubMed intro/conclusion去重过滤后多阶段4096fullarticle→1024MIMICIII+abstract→512abstract预训练；NER tokenhead与RE pooledhead共享encoder、分别跨datasets CE multitaskfinetune。", "evaluation_evidence": "PubMedarticle2.3M/10Btokens、abstract42M/10B、MIMIC2M/2B；三阶段1.89M/1.37M/1.5Msteps。BLURB13tasks与额外NER/RE，Stage3/MTStage3对PubMedBERT/BioLinkBERT/BERN2；并非每task都好，JNLPBA multitask78.55→77.40，QA多项较弱。", "caveats": [ "MIMIC不是可无条件获取公开数据，需credential/DUA，下一批只能已授权或公开PubMed/benchmark子集，不另获患者数据。", "4096longcontext训练动机不能仅以短任务表证明长依赖增益，需长度与multitask预算控制。", "公开collection链接不表示所有训练数据/完整pipeline开放，未独立核作者权重版本。", "仅biomedicaltextNER/RE诊断，不形成临床诊断建议。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://huggingface.co/collections/bisectgroup/biobigbird", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11430v1#S3", "https://arxiv.org/html/2610.11430v1#S4", "https://arxiv.org/html/2610.11430v1#S5", "https://arxiv.org/html/2610.11430v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8562, 8962 ], [ 700, 1700 ], [ 1700, 4800 ], [ 5900, 8000 ] ] }, { "section_index": 3, "heading": "4 Experimental Setup", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 5087, 5487 ] ] }, { "section_index": 4, "heading": "5 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3633, 4033 ], [ 0, 1800 ] ] }, { "section_index": 6, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 190 ], [ 0, 190 ], [ 0, 190 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11430.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11422", "assignment_index": 132, "title": "Causal-fate dynamics of unrealized influence", "published": "2026-10-08T07:48:53Z", "authors": [ "Yiwei Liu", "Luwei Yang", "Shunbo Lei" ], "affiliation": [ "Yiwei Liu Email: yiweiliu1@link.cuhk.edu.cn Affiliation: School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, Guangdong, 518172, China Luwei Yang Email: yangluwei@sribd.cn Affiliation: Shenzhen Research Institute of Big Data (SRIBD), Shenzhen, 518172, China Shunbo Lei Email: leishunbo@cuhk.edu.cn Affiliation: School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, Guangdong, 518172, China Affiliation: Shenzhen Research Institute of Big Data (SRIBD), Shenzhen, 518172, China" ], "source_url": "https://arxiv.org/html/2610.11422v1", "source_sha256": "21e8ea4a1b8053d415589bbd52464be1b52760ca6d00d3449626ab9940ea4c25", "source_bytes": 512876, "source_json_sha256": "a739d099a5986cbba57ad782d019e24ab3ed83ef2c0e0dae35ebda37513596e7", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "causal-fate在realizedH、latentxi、completeZ=H+xi上按fullmapF和tokenlocalmapL构造h=F(Z)-L(H)，threshold每token释放整个h或保latent；Hnext=L(H)+psi，xinext=h-psi保证complete层递推严格fullmap。最终logit只realizedstate，terminallatent不强制释放。另connectome模拟与crossobserverBGPhistory现象为不同证据等级。", "evaluation_evidence": "frozenSmolLM2-135M/WikiText2 length128，33valblocks4191positions/65test8255positions，validation选非trivial最低realizationratio且PPL≤1.06base；testfull38.6512/selective39.9138、ratio.884022。BGP publicRouteViewsChile/RIPErrc06 2025-07-06一天MRT、5,568,057records无parseerror；这是operationalnetworkobservationaldata，不是介入用户AB。", "caveats": [ "完整F(Z)仍每层计算，release ratio/active layer ratio不是attentioncost或computespeed指标，不能宣称推理加速。", "真实虫神经机理只是模型启发hypothesis，routingpredictivehistory只兼容而不识别唯一physicallatent mechanism。", "exacttransport在指定additivecoordinates/maps/reference的构造内成立，不是所有自然系统本体论证明。", "很小WikiTextsubset、无fine-tuning单模型，下一批只能机制诊断且保留PPL退化；不得把terminallatent补释放改算法。" ], "online_ab": "real public operational Internet-routing archives are reviewed, but observational history-state evidence is not randomized production A/B", "upstream_assets": [ { "url": "https://github.com/Hotaru366/causal-fate-code", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/francescorandi/wormneuroatlas", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://github.com/francescorandi/wormneuroatlas.git", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/HuggingFaceTB/SmolLM2-135M", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/Salesforce/wikitext", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11422v1#S2", "https://arxiv.org/html/2610.11422v1#S3", "https://arxiv.org/html/2610.11422v1#S4", "https://arxiv.org/html/2610.11422v1#Sx1", "https://arxiv.org/html/2610.11422v1#Sx2", "https://arxiv.org/html/2610.11422v1#as1_Sx3", "https://arxiv.org/html/2610.11422v1#as1_Sx4" ], "evidence_sections": [ { "section_index": 1, "heading": "2 Results", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 0, 2000 ], [ 13000, 16000 ] ] }, { "section_index": 2, "heading": "3 Discussion", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 2850, 3250 ] ] }, { "section_index": 3, "heading": "4 Methods", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8539, 8939 ] ] }, { "section_index": 4, "heading": "Data availability", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1416, 1816 ], [ 0, 700 ] ] }, { "section_index": 5, "heading": "Code availability", "read_character_ranges_in_cached_section": [ [ 0, 388 ], [ 0, 388 ], [ 0, 388 ] ] }, { "section_index": 11, "heading": "Supplementary Note 3: BGP routing history-state identification", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 0, 1600 ] ] }, { "section_index": 12, "heading": "Supplementary Note 4: Transformer architecture and evaluation protocol", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 3889, 4289 ], [ 0, 2600 ], [ 2600, 4289 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11422.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11389", "assignment_index": 133, "title": "Fact over Fiction: Detection of Pathological Hallucinations in Sinhala-to-English Neural Machine Translation", "published": "2026-10-08T07:19:51Z", "authors": [ "Navam Obeysekara", "Nevidu Jayatilleke" ], "affiliation": [ "Navam Obeysekara Affiliation: School of Computing,\nInformatics Institute of Technology,\nSri Lanka Email: navam.20211058@iit.ac.lk Nevidu Jayatilleke Affiliation: Department of Computer Science & Engineering, University of Moratuwa, Sri Lanka Email: nevidu.25@cse.mrt.ac.lk" ], "source_url": "https://arxiv.org/html/2610.11389v1", "source_sha256": "5d745a06662a7fa1fe83ff07e7f818b19c69224902f3befcc005901f518188ec", "source_bytes": 224631, "source_json_sha256": "1610097215743daabb4fc5de83f6ed9150375568de3174f3835975d9c9ab436a", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "SinhalaEnglishref-free detector mDeBERTa(S,H)tokenbinarylabel；从Englishreference做entity/antonym/dependency/numeric及EnglishreencodehighT的probabilisticchain，difflib标签replace/insert、charrescue.80处理形态变体，source/specialmask。neuralrisk、wordlogprob、LaBSE三signalensemble。", "evaluation_evidence": "24996cleanparallel抽7500source形成45000rows；detectorbalanced15000rows source-disjoint10486train1497val3017test/1485testsource。三seedF1.841±.001，shuffle/removesourceAUROC .970→chance，但hypothesisonlyF1.622；8NMTbenchmark仅detectorfiring51..433不是验证真实hallucinationrate。", "caveats": [ "syntheticEnglishedits/Englishreencode不代表Sinhala-originnaturalerrors，真实generalization未验证。", "sourceablation不能消除monolingualartifact；difflib不认lexicalparaphrase且omission无标签，charrescue可能错误救短词。", "LaBSEagreement不等precision，logprob30%flag为构造阈值，benchmark单seed无humanGT不能排名真实安全性。", "nextbatch需sourcegroup split、validation校准与人工自然错误audit，不能以source/refcorruption提供testgold给translationmodel。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/Imanjith/synthetic-sinhala-english-dataset", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/NLPC-UOM/nllb-top25k-ensi-cleaned", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/datasets/imanjith/synthetic-sinhala-english-hallucination-dataset", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" }, { "url": "https://huggingface.co/thilina/mt5-sinhalese-english", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11389v1#S3", "https://arxiv.org/html/2610.11389v1#S4", "https://arxiv.org/html/2610.11389v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ], [ 2300, 5300 ], [ 6700, 9700 ] ] }, { "section_index": 3, "heading": "4 Evaluation", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8813, 9213 ], [ 4600, 7300 ] ] }, { "section_index": 5, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1216, 1616 ], [ 0, 1500 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11389.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11374", "assignment_index": 134, "title": "Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder", "published": "2026-10-08T07:06:58Z", "authors": [ "Zidan Wang", "Yaqian Li", "Xiaokai Zhang", "Kaiwen Long", "Kun He", "Hanpeng Liu" ], "affiliation": [ "Zidan Wang Affiliation: Huazhong University of Science and Technology Yaqian Li Affiliation: Li Auto Inc. Xiaokai Zhang Affiliation: Huazhong University of Science and Technology Kaiwen Long Affiliation: Li Auto Inc. Kun He Affiliation: Huazhong University of Science and Technology Hanpeng Liu Affiliation: Huazhong University of Science and Technology Affiliation: Li Auto Inc." ], "source_url": "https://arxiv.org/html/2610.11374v1", "source_sha256": "3031139a15f19d315ec4b55a4bd1af636aa97ab1c9301f67a0de6f1e6f048fd1", "source_bytes": 609237, "source_json_sha256": "022109d85ee479faa9269d680cadf6e14d0f0914cdd4b25262d104128515f85b", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "ComCLIP冻结textencoder与DINOv2teacher，只更新CLIPvision单epoch：小tau.01 symmetricInfoNCE集中hardnegatives，postprojectionMSE锚originalCLIP，rowsoftmax selfsimilarity(T.1)的DINO→CLIP KL relationaldistill，不直接把CLIPfeature对齐DINOcoordinate。", "evaluation_evidence": "ViTB16/L14、CC3M/12M/COCO caption、多seed；zero-shot≈CLIPRefine、linearprobeB16 48.99vs42.28、MMVPL14 24.20vs19.01，但retrievalCLIPRefine更强；dropinLLaVA1.5-7B projector/LLM不重align，8benchno netchange。", "caveats": [ "小taugradient论证仅InfoNCE、非KLselfdistill或SigLIPsigmoid；SigLIP实验换成该InfoNCE目标。", "MMVP135pairs高seedvariance，仅L14对Refine显著，不是fine-grainedperceptionSOTA。", "dropincompatibility不等下游VLM能力改善，未经realign的net0必须保留。", "未来需validationtemperature/weights、数据匹配和真实三loss/teacher，不以分类head或syntheticfeatures代理。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [ { "url": "https://github.com/showstarpro/ComCLIP.git", "status": "link in original text; authorship/content/availability not independently verified in this bounded review" } ], "evidence_anchors": [ "https://arxiv.org/html/2610.11374v1#S3", "https://arxiv.org/html/2610.11374v1#S4", "https://arxiv.org/html/2610.11374v1#S5" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Method", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11376, 11776 ], [ 4200, 7500 ], [ 7900, 10300 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 15600, 16000 ] ] }, { "section_index": 4, "heading": "5 Conclusion and Limitations", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 1702, 2102 ], [ 0, 1800 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11374.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11363", "assignment_index": 135, "title": "UniData: Universal Multimodal Instruction Generation Pipeline", "published": "2026-10-08T06:52:24Z", "authors": [ "Jiaqi Tang", "Yi-Feng Wu", "Yuting Zhang", "Hao Lu", "Bowen Fu", "Qing-Guo Chen", "Xiaogang Xu", "Yuwei Hu", "Shiyin Lu", "Wei Wei", "Lei Zhang", "Zhao Xu", "Weihua Luo", "Qifeng Chen", "Ying-Cong Chen" ], "affiliation": [ "Jiaqi Tang † † thanks: Equal contribution. Affiliation: The Hong Kong University of Science and Technology Yi-Feng Wu 1 1 footnotemark: 1 Affiliation: ATH, Alibaba Group Yuting Zhang 1 1 footnotemark: 1 Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Hao Lu Affiliation: The Hong Kong University of Science and Technology (Guangzhou) Bowen Fu Affiliation: Northwestern Polytechnical University Qing-Guo Chen Affiliation: ATH, Alibaba Group Xiaogang Xu Affiliation: Zhejiang University {jtang092, cqf}@ust.hk , yingcongchen@hkust-gz.edu.cn Yuwei Hu Affiliation: ATH, Alibaba Group Shiyin Lu Affiliation: ATH, Alibaba Group Wei Wei Affiliation: Northwestern Polytechnical University Lei Zhang Affiliation: Northwestern Polytechnical University Zhao Xu Affiliation: ATH, Alibaba Group Weihua Luo Affiliation: ATH, Alibaba Group Qifeng Chen † † thanks: Corresponding authors. Affiliation: The Hong Kong University of Science and Technology Ying-Cong Chen 2 2 footnotemark: 2 Affiliation: The Hong Kong University of Science and Technology (Guangzhou)" ], "source_url": "https://arxiv.org/html/2610.11363v1", "source_sha256": "492f8cc1c6599112a8c11312d62c2f8be7e0576a2e62e4211b0df56e1e2552e3", "source_bytes": 366856, "source_json_sha256": "d18a217254e9e55692552056787ef28a44b683f405424f9d81f829a6ae1f8b95", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "UniData以keyword+随机context DEG生成多events，MIG textcenteredLLaMA3将Ovisvisual/Whispermusicembedding及其他六模态textrepresentation整合，预测image/musicprompt而非nativepixel/audio，调用SD3/MusicGen/emoji/search/qrcode实现输出。CIF窗口比较新旧round重复则提高repetitionpenalty、不相关则提高temperature，规则改sampling。", "evaluation_evidence": "UniDataset20000entries/九输出类型，随机90%train10%test、冻结visual/musictokenizer；SelfInstructGPT4/VIGCVicuna基线与GPT-guidedquality、附加userstudy/downstreamfinetune及DEG/CIFablation。", "caveats": [ "九模态是textcenteredtoolsorchestration，不是nativeany-to-anyfoundationmodel；image/music质量受外部generator限。", "随机entrysplit未建立event/source/groupdisjoint，GPTjudge与LLaMA/Vicuna/GPT4backbone不匹配需下一批核公平性。", "提高temperature应对irrelevance是heuristic，不是可靠grounding/verifier，longhorizon只有简单规则。", "searchAPI/outputURL/版权与biasaudit未完成，下一批限定许可本地固定publicassets，不发布未审合成数据为可信benchmark。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11363v1#S3", "https://arxiv.org/html/2610.11363v1#S4", "https://arxiv.org/html/2610.11363v1#S5", "https://arxiv.org/html/2610.11363v1#Sx1" ], "evidence_sections": [ { "section_index": 2, "heading": "3 Dataset Construction", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 4009, 4409 ] ] }, { "section_index": 3, "heading": "4 Methodology", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 11588, 11988 ], [ 1000, 4100 ], [ 4100, 7800 ], [ 7800, 10400 ] ] }, { "section_index": 4, "heading": "5 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 700 ], [ 8609, 9009 ], [ 0, 1700 ] ] }, { "section_index": 6, "heading": "Limitations", "read_character_ranges_in_cached_section": [ [ 0, 566 ], [ 166, 566 ], [ 0, 566 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11363.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11329", "assignment_index": 137, "title": "PathLang: A Language-Centered Benchmark for Vision-Language Models in Computational Pathology", "published": "2026-10-08T06:27:42Z", "authors": [ "Fanqi Cheng", "Kuo Gong", "Shangke Liu", "Beidi Zhao", "Junchao Zhu", "Zheyu Zhu", "Leiyue Zhao", "Fengbei Liu", "John Cannon", "Gang Wang", "Zu-hua Gao", "Kenji Ikemura", "Yihe Yang", "Yaohong Wang", "Yuankai Huo", "Xiaoxiao Li", "Mert R. Sabuncu", "Ruining Deng" ], "affiliation": [ "Fanqi Cheng Address: New York University, New York, NY, USA Kuo Gong Address: Columbia University, New York, NY, USA Shangke Liu Address: Weill Cornell Medicine, New York, NY, USA Address: Cornell University, Ithaca, NY, USA Beidi Zhao Address: University of British Columbia, Vancouver, BC, Canada Junchao Zhu Address: Vanderbilt University, Nashville, TN, USA Zheyu Zhu Address: University of Pennsylvania, Philadelphia, PA, USA Leiyue Zhao Address: Johns Hopkins University, Baltimore, MD, USA Fengbei Liu Address: Cornell University, Ithaca, NY, USA John Cannon Address: New York Medical College, New York, NY, USA Gang Wang Address: University of British Columbia, Vancouver, BC, Canada Address: BC Cancer Agency, Vancouver, BC, Canada Zu-hua Gao Address: University of British Columbia, Vancouver, BC, Canada Address: BC Cancer Agency, Vancouver, BC, Canada Kenji Ikemura Address: Weill Cornell Medicine, New York, NY, USA Yihe Yang Address: Weill Cornell Medicine, New York, NY, USA Yaohong Wang Address: The University of Texas MD Anderson Cancer Center, Houston, TX, USA Yuankai Huo Address: Vanderbilt University, Nashville, TN, USA Xiaoxiao Li Address: University of British Columbia, Vancouver, BC, Canada Mert R. Sabuncu Address: Weill Cornell Medicine, New York, NY, USA Address: Cornell Tech, New York, NY, USA Ruining Deng Email: rud4004@med.cornell.edu Corresponding author: Corresponding author. Address: Weill Cornell Medicine, New York, NY, USA" ], "source_url": "https://arxiv.org/html/2610.11329v1", "source_sha256": "aebbd58b50f92b093f0eb4a3d55baf825c3490b771d7ee0077bcc565274ac580", "source_bytes": 1388352, "source_json_sha256": "5f08b865787e82a5a0a3910aad55bd9ac3b54663ce30036d6d1287b0b2ee35a8", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "PathLang冻结slides/labels及image-textdirection，只变临床diagnosticphrase：zero-shot分类、crossmodalretrieval、semanticallyequivalent/length/style paraphrase、promptensemble、四竞争candidatepool。九patchdualencoders queryconditionedtopKtilepool统一slideinterface，并对四ratio和meanpool敏感性分析，六boardcertifiedpathologist审prompt/pool。", "evaluation_evidence": "PANDA10609/UniToPatho237/CAMELYON16 399/CAMELYON17 499/TCGAGBMLGG1335、四organ；临床等义改写高predictioninstability，meanpool仍存在。主表每model/dataset(/pool)按evaluationlabel选最好aggregation，T3同配置固定全部variant/ensembles，所以只是oracleperformance reference。", "caveats": [ "oracleaggregation不是validation-selecteddeployable能力比较，不得把testgold用于实际模型选择；nextbatch需固定meanpool或validation-only选r。", "不同crossmodalretrieval方向candidate数不同，只descriptive，T2positivephrases不等T1matchedprompt。", "Set2/3poolcomposition同时变化，不是控制sameorgan/crossorgan单轴因果。", "publicWSI许可、patient/sitegroup隔离还需核，只做去标识公开diagnostic，非患者临床诊断。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11329v1#S3", "https://arxiv.org/html/2610.11329v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 PathLang: Language-Centered Benchmarking", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 10434, 10934 ], [ 1000, 2900 ], [ 6500, 10934 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 15500, 16000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11329.html", "source_is_research_input_not_authority": true }, { "arxiv_id": "2610.11317", "assignment_index": 138, "title": "DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition", "published": "2026-10-08T06:23:01Z", "authors": [ "Yuxuan Lou", "Kai Yang", "Geng Zhang", "Yong Liu", "Yang You" ], "affiliation": [ "Yuxuan Lou † † thanks: Equal contribution. Affiliation: School of Computing, National University of Singapore, Singapore Kai Yang 1 1 footnotemark: 1 Affiliation: Department of Computer Science and EngineeringShanghai Jiao Tong University, Shanghai, China Geng Zhang Affiliation: School of Computing, National University of Singapore, Singapore Yong Liu Affiliation: School of Computing, National University of Singapore, Singapore Yang You † † thanks: Corresponding author: youy@comp.nus.edu.sg . Affiliation: School of Computing, National University of Singapore, Singapore" ], "source_url": "https://arxiv.org/html/2610.11317v1", "source_sha256": "6ff361abbb5a24b9fff435cc3a67a5fdd33dc062663411d5a3bc1eb3bfb3b031", "source_bytes": 592844, "source_json_sha256": "f4d2af97f9c650d94b1fc2c71fca98ae1849ea193d9923458b0a0a265f42b9a7", "status": "reviewed_deferred", "qualifies_for_reproduction_consideration": true, "selection_priority": "manual method/evaluation review; not keyword autoaccept", "in_frozen_16_paper_batch": false, "next_batch": true, "implementation_status": "not_implemented", "full_paper_read": false, "read_scope": "abstract + explicitly bounded original method/evaluation passages (ranges recorded); not complete main sections or appendices", "defining_method": "DivMoE从同base进行math/code/science/common四domainfullsequenceCPT，再每SwiGLUFFN gate/up按row、down按column配对切两shards为八fineexperts，magnitudecorrectiongamma=(E*G²/k)^(1/3)，nonexpertweights取base+specialistsmean。router各domain选top1，再跨group选topk=2、hard atmostone/group，softmaxweight仍非equal，随后统一Stage2CPT/Stage3SFT。", "evaluation_evidence": "Qwen3-1.7BBase/Llama3.2-1B及4B→12B，15bench六upcyclingbaseline；Stage2所有500B tokens，BTX/DivMoE另Stage1共40B、总540B。Qwen1.7平均55.6vsbestbaseline51.6，Dropfine23.2nearfromscratch22.2；publicTulu3SFT同应用matchedvariants，12B全8bench64.5=Moonlight16B。", "caveats": [ "hardcross-domainselection只保证不同group，不保证global专家负载均衡/语义diversity或能力提升。", "denseequivalent用近似domainadaptation假设，split与magnitudecorrection/frozeninitialweights需真实检查，不能用randomdistinctexperts代理。", "Stage2budgetmatched不等pipelinecompute完全matched：DivMoE/BTX需额外40B specialistCPT，Moonlightcomparison并非同预训练语料/compute。", "下一批完整540B成本超本批，boundedconversion/routing机制可L1，正式能力需公开实际训练与isolatedvalidation/test，不宣称已实现。" ], "online_ab": "not established from these bounded original passages; not an abstract-only rejection", "upstream_assets": [], "evidence_anchors": [ "https://arxiv.org/html/2610.11317v1#S3", "https://arxiv.org/html/2610.11317v1#S4" ], "evidence_sections": [ { "section_index": 2, "heading": "3 DivMoE", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 10036, 10536 ], [ 900, 4200 ], [ 4200, 7400 ] ] }, { "section_index": 3, "heading": "4 Experiments", "read_character_ranges_in_cached_section": [ [ 0, 900 ], [ 7066, 7566 ], [ 900, 2000 ] ] } ], "source_artifact": ".cache/oct12-scan/fulltext/2610.11317.html", "source_is_research_input_not_authority": true } ] }