{ "schema_version": 1, "batch": "2026-10-26-priority-components", "delivery_status": "bounded_scope_complete", "selected_scope_complete": true, "all_sources_exhaustive": false, "review_watermark_advanced": false, "review_watermark": { "before": "2026-10-15", "after": "2026-10-15" }, "counts": { "selected_papers": 16, "components_accepted": 8, "components_pending_verification": 0, "gpu_validated_components": 8, "reference_only_high_cost_done": 8, "capability_reproductions_claimed": 0, "category_entries_including_crosslists_and_replacements": 1180, "unique_bulletin_identities": 835, "snapshot_added_base_identities_not_new_paper_count": 802, "snapshot_added_version_strings_not_revision_count": 832, "section_entry_counts": { "new": 446, "cross-list": 289, "replacement": 445 }, "selected_first_publications_after_previous_capture": 0 }, "fixed_component_keys": [ "detector-scope", "dt-jepa", "editjev", "frontier-opd", "hge", "mart-grpo", "opst", "uniaction" ], "reference_only_keys": [ "flexglue", "harnessmm", "image-rsi", "lazyflash", "paperorchestra", "pathjev", "rsi-opt", "watch-and-learn" ], "papers": [ { "id": "2610.30006", "key": "detector-scope", "domain": "foundation-models", "title": "Building Auditable AI Text Detectors with Sparse Autoencoders", "priority": "P0", "paper_url": "https://arxiv.org/abs/2610.30006v1", "detail_path": "foundation-models/2610.30006-detector-scope/README.md", "first_author": "Liam Dugan", "first_author_affiliation": "University of Pennsylvania", "original_published_at": "2026-10-21T16:41:14Z", "decision": "component_only", "intake_class": "recent-source-release-backfill", "first_publication_after_baseline_capture": false, "observed_versions": [], "bulletin_categories": [], "bulletin_entry_kinds": [], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Section4、Section5及AppendixA.3/B.1/B.2；https://arxiv.org/html/2610.30006v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "Gemma2-2B、四层各16384维SAE、COLING训练610k文档及PRDB；作者报告SAE训练185 B200 GPU小时、neural baseline训练80 A6000 GPU小时，另Gemini3.1FlashLite特征解释/验证和多条件实验。", "estimated_cost_basis": "GPU小时为作者报告，不是本地测量；完整多条件训练和大规模特征解释超单卡短时预算。小张量head诊断无需模型、语料或API。", "decision": "完整实验成本过高，不复现（审查 DONE）；不下载权重/语料、不试跑、不自动重试。仅独立交付可执行L1稀疏检测头及归因机制。" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/foundation-models/2610.30006-detector-scope/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32657", "key": "dt-jepa", "domain": "foundation-models", "title": "DT-JEPA: Learning to Predict Future Representations for Decision-Making", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32657v1", "detail_path": "foundation-models/2610.32657-dt-jepa/README.md", "first_author": "Changxu Ni", "first_author_affiliation": "Southeast University", "original_published_at": "2026-10-23T14:04:18Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32657v1" ], "bulletin_categories": [ "cs.LG" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1正文/实验/附录：https://arxiv.org/html/2610.32657v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "D4RL共19个locomotion/navigation数据配置、主表5种子、多个未见reward的在线适配及基线/消融；附录DT4层/4头/512hidden、batch256/context20。原文未给总GPU小时，不臆测。完整offline+online矩阵超单卡短时预算，不安装MuJoCo、不下载轨迹、不启动环境试跑。", "estimated_cost_basis": "原文配置与矩阵的成本审查，不是本地完整实验测量；无需启动大模型/数据来试探成本。", "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。只保留独立L1机制组件，不作为模型能力或evolve晋级证据。" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/foundation-models/2610.32657-dt-jepa/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32635", "key": "editjev", "domain": "foundation-models", "title": "EditJev: System One Routing for Image Editing Models", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32635v1", "detail_path": "foundation-models/2610.32635-editjev/README.md", "first_author": "Zhile Guan", "first_author_affiliation": "Peking University", "original_published_at": "2026-10-23T13:46:53Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32635v1" ], "bulletin_categories": [ "cs.AI", "cs.CV" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1正文/实验/附录：https://arxiv.org/html/2610.32635v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "完整方案Qwen3-VL-4B LoRA router，五个编辑器FLUX.2 Klein4B/9B、Kontext、OmniGen2、Qwen-Image-Edit（20B），离线所有候选生成及Qwen3-VL-32B/GPT judge，737 ImgEdit+606 GEdit-En请求和多组消融。论文RTX PRO6000延迟计时不适合伪装成A100本地实测，不下载多套编辑权重或批量调API。", "estimated_cost_basis": "原文配置与矩阵的成本审查，不是本地完整实验测量；无需启动大模型/数据来试探成本。", "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。只保留独立L1机制组件，不作为模型能力或evolve晋级证据。" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/foundation-models/2610.32635-editjev/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32976", "key": "flexglue", "domain": "multimodal-models", "title": "FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation", "priority": "P0", "paper_url": "https://arxiv.org/abs/2610.32976v1", "detail_path": "multimodal-models/2610.32976-flexglue/README.md", "first_author": "Philipp Lindenberger", "first_author_affiliation": "ETH Zurich", "original_published_at": "2026-10-23T17:59:00Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32976v1" ], "bulletin_categories": [ "cs.CV" ], "bulletin_entry_kinds": [ "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 PDF第7页Implementation Details及全文实验", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "冻结DINOv3 ConvNeXt-B，9层512维Transformer，8数据集，50epochs×150Ksamples，4H200约4GPU-days；gradient checkpointing后4RTX4090约1.5days，另10epochs disambiguation fine-tune及多benchmark/SfM。", "estimated_cost_basis": "严格引用论文配置与时间口径；不是本地实测，不将GPU-days误写成单卡可完成的几小时。", "decision": "成本过高，不复现（审查 DONE）；完整模型/kernel/多数据集训练和评测均不下载、不试跑、不自动重试。" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32802", "key": "frontier-opd", "domain": "agent-research", "title": "Frontier-OPD: Discovering the Edge of Agent Capability and Teaching Agents to Go Beyond", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32802v1", "detail_path": "agent-research/2610.32802-frontier-opd/README.md", "first_author": "Zheng Jiang", "first_author_affiliation": "Tsinghua University", "original_published_at": "2026-10-23T15:40:44Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32802v1" ], "bulletin_categories": [ "cs.CL" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "v1 Sec3-4, Appendix A-C, Fig4 efficiency study", "model_scale": "Qwen3-32B or Qwen3-8B-RL teacher; Qwen3-1.7B/4B student", "paper_configuration": "ALFWorld/WebShop/ScienceWorld;200 training steps; M4 B4 K1 H3; prompt10240/response512;3seeds; efficiency study uses4 NVIDIA B300", "available_resource": "single NVIDIA A100 or A30; local Mac", "decision": "成本过高，不复现（审查 DONE）；不下载模型、不启动真实环境矩阵、不自动重试。仅L1选择和验证协议" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/agent-research/2610.32802-frontier-opd/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32813", "key": "harnessmm", "domain": "agent-research", "title": "HarnessMM: Recursive Self-Improvement for Harness Token Efficiency", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32813v1", "detail_path": "agent-research/2610.32813-harnessmm/README.md", "first_author": "Yuhua Jiang", "first_author_affiliation": "Tsinghua University", "original_published_at": "2026-10-23T15:47:47Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32813v1" ], "bulletin_categories": [ "cs.AI" ], "bulletin_entry_kinds": [ "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Sec3-5.1、AppendixC/D/L", "available_resource": "单张A100/A30或本地Mac", "configuration": "真实Codex/MiMo/DeepSeek原生harness及HumanEval/MBPP等完整taskset，配合TRAIN/validation/TEST和跨领域QA/ALFWorld/AIME；完整计入作者、selector、fallback、失败与cacheinput请求", "decision": "成本过高，不复现（审查 DONE）；不启动大模型/API/原生harness实验矩阵，不下载模型，不自动重试" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32420", "key": "hge", "domain": "recommendation", "title": "Hierarchical GTV Estimation: Bridging the Gap between Container-Level Ranking and Item-Level Conversion", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32420v1", "detail_path": "recommendation-components/2610.32420-hge/README.md", "first_author": "Zerong Lan", "first_author_affiliation": "Meituan", "original_published_at": "2026-10-23T09:50:06Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32420v1" ], "bulletin_categories": [ "cs.IR" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source_url": "https://arxiv.org/html/2610.32420v1", "source_location": "Sections4.1--4.2", "data_scale": "712M private industrial records, 121M users, 949K POIs; authors list683M train/29M test", "paper_configuration": "K10, categorical embedding8, HSAE256, towers1024/512/256 Dice, DIN+DCN regression, batch6000, Adam9e-4, one epoch; online A/B13days. No reported device or GPU hours", "available_resource": "One NVIDIA A100 or A30, or local Mac", "estimated_cost_basis": "The private683M-record training and online traffic experiment exceed the affordable single-device scope; cost inference from author configuration, not local timing", "decision": "成本过高，不复现（审查 DONE）；只执行外部embedding/price/regression输入上的HSAE、PLL与RCEF张量和梯度诊断，不下载生产数据、不执行线上实验、不自动重试" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/recommendation-components/2610.32420-hge/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32120", "key": "image-rsi", "domain": "agent-research", "title": "ImageRSI: Verification-Driven Recursive Self-Improvement for Agentic Image Generation", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32120v1", "detail_path": "agent-research/2610.32120-image-rsi/README.md", "first_author": "Shuo Han", "first_author_affiliation": "University of Science and Technology of China", "original_published_at": "2026-10-23T03:42:21Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32120v1" ], "bulletin_categories": [ "cs.AI", "cs.CV" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Sec3/4.1与AppendixA/D", "available_resource": "单张A100/A30或本地Mac", "configuration": "Qwen-Image2/Wan2.7/NanoBanana生成；MindBench500请求、ImageReq431test/530experience、约2000judge训练请求；每题最多3生成轮，Gemini3.1Pro外部评测及SFT/GRPO训练", "decision": "成本过高，不复现（审查 DONE）；不下载权重/数据、不启动API/模型矩阵、不自动重试" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.31911", "key": "lazyflash", "domain": "foundation-models", "title": "LazyFlash: Forward-Dense, Backward-Sparse Attention for Accelerated Long-Context Fine-Tuning", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.31911v1", "detail_path": "foundation-models/2610.31911-lazyflash/README.md", "first_author": "Keqi Deng", "first_author_affiliation": "Microsoft", "original_published_at": "2026-10-22T21:04:07Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.31911v1" ], "bulletin_categories": [ "cs.CL" ], "bulletin_entry_kinds": [ "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1正文/实验/附录：https://arxiv.org/html/2610.31911v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "原文8张H200、65,536长度、2500steps，多个Qwen3模型和语音50K小时训练数据；所有计时在H200。自定义FlashAttention反向kernel及跨长度/模型/精度验证本身也超过本轮单卡低成本范围。不安装/编译kernel或加载权重来“试试”。", "estimated_cost_basis": "原文配置与矩阵的成本审查，不是本地完整实验测量；无需启动大模型/数据来试探成本。", "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。无本地指标/收据，不以通用代理或空协议冒充方法。" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32431", "key": "mart-grpo", "domain": "post-training", "title": "When to Stop Penalizing Actions? Martingale GRPO for VLA Reinforcement Learning", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32431v1", "detail_path": "post-training/2610.32431-mart-grpo/README.md", "first_author": "Zongyao Zhao", "first_author_affiliation": "The University of Hong Kong", "original_published_at": "2026-10-23T09:56:27Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32431v1" ], "bulletin_categories": [ "cs.AI" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "v1 Sec3, Appendix D Table3 and Appendix E", "model_scale": "OpenVLA-OFT actor and frozen Qwen3.8-27B VLM with LoRA event adapter", "paper_configuration": "8 NVIDIA H100 server capacity;7 RoboTwin2 tasks;512trajectories/update;40updates;128seen+128held-out episodes every5updates;1024failed videos plus success references for event model", "available_resource": "single NVIDIA A100 or A30; local Mac", "decision": "成本过高，不复现（审查 DONE）；不下载权重、不启动机器人/VLM完整矩阵、不自动重试。只验证L1事件映射和损失" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/post-training/2610.32431-mart-grpo/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2610.32770", "key": "opst", "domain": "post-training", "title": "On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models", "priority": "P0", "paper_url": "https://arxiv.org/abs/2610.32770v1", "detail_path": "post-training/2610.32770-opst/README.md", "first_author": "Leon Sick", "first_author_affiliation": "Google", "original_published_at": "2026-10-23T15:12:03Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32770v1" ], "bulletin_categories": [ "cs.AI", "cs.CL", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "v1 Sec3-4, Appendix A2/A5, PDF p5 Eq5-6", "model_scale": "Gemma4 E2B/E4B families; also Qwen3-4B-Base and visual extension. E2B/E4B labels are not asserted to be exact total parameter counts", "paper_configuration": "Default7 self-training rounds,285 generated samples/round, repeated fresh-context/paraphrase verification and G student rollouts, student/EMA teacher scoring, multiple domains/thinking settings/baselines", "available_resource": "single NVIDIA A100 or A30; local Mac", "decision": "成本过高，不复现（审查 DONE）；不下载权重、不试跑完整课程，不自动重试。canonical JSD歧义单列，不由成本标签掩盖" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/post-training/2610.32770-opst/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2604.05018", "key": "paperorchestra", "domain": "agent-research", "title": "PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing", "priority": "P0", "paper_url": "https://arxiv.org/abs/2604.05018v1", "detail_path": "agent-research/2604.05018-paperorchestra/README.md", "first_author": "Yiwen Song", "first_author_affiliation": "Google Cloud AI Research", "original_published_at": "2026-04-06T18:00:00Z", "decision": "not_reproduced_high_cost", "intake_class": "historical-directory-backfill", "first_publication_after_baseline_capture": false, "observed_versions": [], "bulletin_categories": [], "bulletin_entry_kinds": [], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Section4/5、AppendixB/D；https://arxiv.org/html/2604.05018v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac；不假设免费无限外部API", "paper_configuration": "Gemini3.1Pro写作、Gemini3Flash联网检索、图像模型/PaperBanana、三轮refinement；作者报告单文60–70次LLM调用、39.6分钟裁尾均值；200篇PaperWritingBench和多judge/多基线及11位研究者评测。", "estimated_cost_basis": "调用量/均值为作者在10篇子集上的报告，不是本地测量；按完整评测矩阵需大量联网API调用与人工审查，单次fixture不能忠实代表。", "decision": "成本过高，不复现（审查 DONE）；不调用付费API/启动长写作评测、不自动重试；保留方法、泄漏风险、来源和成本记录。" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32478", "key": "pathjev", "domain": "foundation-models", "title": "PathJev: Decision-Native Vision–Language Modelling for Histopathology", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32478v1", "detail_path": "foundation-models/2610.32478-pathjev/README.md", "first_author": "Cristian Tommasino", "first_author_affiliation": "University of Naples Federico II", "original_published_at": "2026-10-23T10:47:40Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32478v1" ], "bulletin_categories": [ "cs.CV", "cs.LG" ], "bulletin_entry_kinds": [ "cross-list", "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1正文/实验/附录：https://arxiv.org/html/2610.32478v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "MedGemma1.5-4B NF4、896×896图像、LoRA+head约39.6M参数、40Kstates；Table7单GPU每seed PathJev全程10.4–10.5h、Gen-SFT16.3–16.5h，另有超参选择/3seeds/多数据集推理、whole-slide。远超单卡短时预算，不下载公开权重来重复验证成本。", "estimated_cost_basis": "原文配置与矩阵的成本审查，不是本地完整实验测量；无需启动大模型/数据来试探成本。", "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。无本地指标/收据，不以通用代理或空协议冒充方法。" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32026", "key": "rsi-opt", "domain": "agent-research", "title": "RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32026v1", "detail_path": "agent-research/2610.32026-rsi-opt/README.md", "first_author": "Binghang Lu", "first_author_affiliation": "Purdue University", "original_published_at": "2026-10-23T01:05:11Z", "decision": "not_reproduced_high_cost", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32026v1" ], "bulletin_categories": [ "cs.AI" ], "bulletin_entry_kinds": [ "new" ], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Sec3/4/5.8与AppendixA", "available_resource": "单张A100/A30或本地Mac", "configuration": "9单GPU训练任务，20agentdecision/run；static5seeds、evolution2seeds+top3重评2seeds、freshfinal3seeds；Opus5.5/GPT5.5/GPT6.1。RSI单run平均3.65–4.61M input与24–42K output token，另有实际模型训练/CPU代理实验", "decision": "成本过高，不复现（审查 DONE）；不运行模型/付费API搜索矩阵，不执行生成代码，不自动重试" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" }, { "id": "2610.32977", "key": "uniaction", "domain": "multimodal-models", "title": "UniAction: Unified Action Conditioning for World Models", "priority": "P1", "paper_url": "https://arxiv.org/abs/2610.32977v1", "detail_path": "multimodal-models/2610.32977-uniaction/README.md", "first_author": "Bowen Xue", "first_author_affiliation": "Stanford University", "original_published_at": "2026-10-23T17:59:13Z", "decision": "component_only", "intake_class": "new-announcement-selected", "first_publication_after_baseline_capture": false, "observed_versions": [ "2610.32977v1" ], "bulletin_categories": [ "cs.CV" ], "bulletin_entry_kinds": [ "new" ], "status": "implemented", "implementation_status": "component_only", "component_implemented": true, "component_accepted": true, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1正文/实验/附录：https://arxiv.org/html/2610.32977v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "Wan2.2-I2V-A14B high-noise expert LoRA，GoalForce/SolarWM/EgoDex/DROID各10K clips，共40K段81帧480×832；410反事实生成、多个motion/pose估计器、4个专家和100人评测。完整视频训练/生成显著超过单A100/A30短时资源，不下载14B权重或40K视频，不试跑成本。", "estimated_cost_basis": "原文配置与矩阵的成本审查，不是本地完整实验测量；无需启动大模型/数据来试探成本。", "decision": "成本过高，不复现（审查 DONE）；完整实验不下载、不试跑、不自动重试。只保留独立L1机制组件，不作为模型能力或evolve晋级证据。" }, "requires_gpu_validation": true, "gpu_validation_passed": true, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "bounded_l1_component_accepted", "component_evidence": "docs/multimodal-models/2610.32977-uniaction/metrics/component-cuda-seeds42-44.json", "reason": "限定组件三种子CPU/真实A100已验收；完整高成本实验不复现、审查DONE，不声称能力或全窗口穷尽。" }, { "id": "2510.04673", "key": "watch-and-learn", "domain": "agent-research", "title": "Watch and Learn: Learning to Use Computers from Online Videos", "priority": "P0", "paper_url": "https://arxiv.org/abs/2510.04673v1", "detail_path": "agent-research/2510.04673-watch-and-learn/README.md", "first_author": "Chan Hee Song", "first_author_affiliation": "The Ohio State University", "original_published_at": "2025-10-06T10:29:00Z", "decision": "not_reproduced_high_cost", "intake_class": "historical-directory-backfill", "first_publication_after_baseline_capture": false, "observed_versions": [], "bulletin_categories": [], "bulletin_entry_kinds": [], "status": "deferred", "implementation_status": "deferred_high_cost", "component_implemented": false, "component_accepted": false, "review_done": true, "full_experiment_review_done": true, "full_experiment_status": "not_reproduced_high_cost", "cost_assessment": { "source": "固定v1 Section3.1/4及AppendixD.3；https://arxiv.org/html/2510.04673v1", "available_resource": "单张NVIDIA A100或A30，或本地Mac", "paper_configuration": "超过630k截图状态转移；SigLIP2+四层Transformer+多动作/坐标分类与GPT2文本头，8张80GB A100训练15epoch；53,125视频轨迹，另7B CUA SFT及OSWorld真环境/API长轨迹。", "estimated_cost_basis": "8A100/15epoch为原文配置，不是本地测量；重建训练语料、IDM、视频标注与多系统真环境评测超预算，且未核验作者代码/权重。", "decision": "成本过高，不复现（审查 DONE）；不下载训练资源、不搭GUI环境、不以规则IDM或gold动作替代核心，不自动重试。" }, "requires_gpu_validation": false, "gpu_validation_passed": false, "automatic_retry": false, "automatic_implementation_queue": false, "promotion_eligible": false, "evolve_integrated": false, "resolution": "not_reproduced_high_cost", "component_evidence": null, "reason": "完整实验成本过高，不复现（审查 DONE）；仅参考资料，无执行入口或自动重试。" } ], "reference_scope_expansion": { "added_reference_only_keys": [ "harnessmm", "image-rsi", "rsi-opt" ], "component_scope_unchanged": true, "existing_acceptance_preserved": true } }