31 条审后延期线索：不等于已接受实施¶
原脚本仅凭‘待实现/下一批须/保留下一批’等自由文本，将这些 reviewed_deferred 项升为待交付。
这里保留全部线索和理由；不将它们假装已完成，也不混进结构化接受的 504 项。
| 论文 | 理由类型 | 原始理由 |
|---|---|---|
| 2610.13166 Scaling Agentic Data for Long-Horizon Terminal Intelligence | integrity_or_acceptance_audit_preconditions | 已读TerminalHorizon方法/实验：独立execution发现合成环境预先完成的关键工作，再reconstruct并扩展同goal任务；1.5k环境、Qwen3.5-27B SFT与12bench。production settings是任务评测不证明线上A/B；下一批须隔离environment gold verifier与agent可见包。 |
| 2610.13144 Competence-Aware Orchestration for LLM-Based Multi-Agent Systems | implementation_or_execution_preconditions | 已读competence orchestration方法/实验：固定probe capability fingerprint加query-conditioned模型，learned role activation与communication graph；6 reasoning/codebench及unseen backbones。下一批须训练真实competence predictor，probe/test隔离而非gold推断能力。 |
| 2610.12957 JoyInside-VLM-Agent: Product-Conditioned Multimodal Policy Learning for Embodied Agents | implementation_or_execution_preconditions | 已读JoyInside方法/实验：产品条件multimodal trajectory、SFT、多teacher OPD及DPER token-level privileged evidence路由；部署单student不读teacher reference。私有120k混合数据、领域skill及两机器人case限制公开复现，下一批须保留真实多阶段训练。 |
| 2610.12794 Meta-Reflective Evolution Effectively Improves Production AI Systems | further_evidence_verification | 已读CoCoEvolve方法与实验协议：archive parent、history检索、live target probing、自测mutation再external评分；train实例可见，validation仅score、test隔离。production AI目标及公开heldout基准不自动证明生产流量A/B，下一批须进一步核验live authority和真实evaluator，非registry标签。 |
| 2610.12655 CLEAT: Learning Value-Guided Clarification for Interactive Agents | implementation_or_execution_preconditions | 已读CLEAT摘要、intent alignment及实验：同history ask/proceed paired rollout估residual query value，controller与task policy co-train，新policy outcome更新labels；Qwen3-4B四bench十设置三seed。下一批须训练真实value predictor+policy，不能拿测试latent goal作agent输入。 |
| 2610.12611 SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments | implementation_or_execution_preconditions | 已读SoftwareGym方法/实验：合成真实可执行GUI app、固定v1 GUI测试迭代repair，182app12类120task，Qwen3.5-9B multimodal SFT、out-of-domain OSWorld。下一批须VM实际GUI交互与合成环境自测隔离，不能用静态截图/固定动作fixture。 |
| 2610.12248 EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams | implementation_or_execution_preconditions | EgoVoice用HoloAssist source-separated/resynth audio和listen-or-speak SFT/DPO训练主动语音助手，170 sessions onset/content/end评测；多模态/语音后训练下一批须实际流式音视频与timing policy。 |
| 2610.11862 LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs | implementation_or_execution_preconditions | 已读LEVER AND/OR graph与实验：partial proof realized objective加open subgoal预测，programmable cost/length/impurity，Lean kernel保correct；matched model/version/price。下一批须真实Lean搜索、预测前评分与cost含失败，不能gold proof排序代替。 |
| 2610.11578 Chronos Enables Code Agents to Reason over Software Evolution | implementation_or_execution_preconditions | 已读Chronos方法/设置：历史PR蒸馏experience cards+typed graph、semantic entry+multi-hop、双patch生成steward选择；6backboneSWE对比。下一批须截止目标issue前history防答案PR泄漏，真实patch执行而非检索静态gold。 |
| 2610.11573 Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies | implementation_or_execution_preconditions | 已读MemoType/TriMEM方法和评测：learned router识别memory/query类型选择retrieval，Contriever统一backbone、LongMemEval/LoCoMo/PerLTQA。下一批须实际router训练与检索标签隔离，人工分类型heuristic不能称核心复现。 |
| 2610.11384 Environmental Feedback Modeling Matters: Rethinking Feedback Treatment in Agentic Hindsight Self-Distillation | implementation_or_execution_preconditions | 已读SELF方法/实验：on-policy真实environment反馈建feedback-conditioned hindsight teacher，joint response modeling loss+selfdistillation；tau/AppWorld reward-free agent训练。下一批须真实双objective梯度，privileged未来feedback仅训练teacher不进部署policy。 |
| 2610.11226 When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization | implementation_or_execution_preconditions | DRQ固定quant grid只refine integer codes，worst-case activation distributions纠正calibration reconstruction泛化，六PTQ methods dense/MoE公开ppl/tasks；下一批须真实整数码/量化forward/uncertainty inner problem。 |
| 2610.11136 The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection | integrity_or_acceptance_audit_preconditions | 已读MARS-Gov方法/实验：legal retrieval、open-set scout动态juror、conservative routing和rewrite语义verification，DGDB/LOCO评测。下一批须独立evidence与重写含义audit，法律背景检索不可视为法律建议/无偏性保证。 |
| 2610.11123 When Interfaces Speak: Data-Aware Generative UI Harness for Active Interaction | implementation_or_execution_preconditions | 已读GenUI task定义/实验：ToolAgent与GUICoder、DynamicUX实际sandbox交互reward、RewardAuditor meta-rubric，UI-TAU真实publicdatabase模拟任务与4B训练。下一批须真实UI/code执行和用户输入，judge reward hacking校验；rounds问卷不当生产A/B。 |
| 2610.11008 Why LLM Agents Favor Their Group: Stakes, Observed Norms, and Reputation | reference_only | 已读群体favoritism实验/结果：任意group、costless/stake、scripted history与个体reputation控制，4400模拟societies。属于agent社会行为测量，保留下一批参考；不能把合成群体pointallocation解释为现实组织因果。 |
| 2610.10786 Plan-and-Patch: Diffusion Language Models for Agentic Planning | implementation_or_execution_preconditions | 已读Plan-and-Patch方法/实验：真实diffusion planner parallelunmask、region infill保prefix/suffix，与独立executor失败region选择，NaturalPlan/ALFWorld/TextCraft训练。下一批须dLLM与联合generation/repair训练，AR字符串编辑非定义机制。 |
| 2610.10305 How Do Transformers Learn to Represent Symmetries? | explicit_unimplemented | Transformerfiniteaugmentation学习pointcloud群不变性与equivariance，受控architecture/ablation；真实训练/OOD几何证据非fixture待实现。 |
| 2610.10120 RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation | implementation_or_execution_preconditions | 已读RewardWeaver method/experiment：admitted rubric语义固定，阶段训练后lowoutcome backwardattribution动态reward allocation，独立controlled expansion；社会/议价/sales。下一批须实际policy训练、validatedrubric稳定，不能追分随意重定义reward或测试goal泄漏。 |
| 2610.10071 HGP:An on-device personalized agent memory via hybrid graph storage | implementation_or_execution_preconditions | 已读HGP方法/实验：confidence-driven LLMannotation在线训lightweightrouter、episode/semantic/proceduralgraphs+workingstate；LoCoMo/PAL选择与QA。下一批须真实router更新与图检索、on-device latency/资源实际测，不能只标签分类fixture。 |
| 2610.10062 Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents | implementation_or_execution_preconditions | 已读loud/quietfailure method/results：自有BFCL statefultool layer四typedfault控制注入，24task1920trial、faultfree nondeterminism基线与检测/改plan/recovery分开。下一批须真实tool状态execute，gold finalstate仅scorer，reasoning无recovery提升负结果保留。 |
| 2610.09952 MSU Team at the Explainable Deepfake Detection Challenge 2026: Grounded Artifact Evidence for Deepfake Detection | explicit_unimplemented | XPlainVerse DINO/Mesorchdetector、trainingartifactmask和patchcontrastive+QwenVL解释/GRPO简化；publictrainval/test可验，真多stage待实现。 |
| 2610.09935 AgentTracer: Tracing Indirect Prompt Injection Attack through Fine-Grained Intention-Execution Alignment | implementation_or_execution_preconditions | 已读AgentTracer graphdesign/evaluation：userintent authorization与operationKB、taskintent implicitdependency、从anomalouscall backwardprune追injectionsource；successattacklogs加背景noise。下一批须完整trace与无attackfalsepositive控制，不能已知injection位置当attributor输入。 |
| 2610.09856 Self-Evolve With a Reference:Anchored Training of Tool-Integrated Agents | implementation_or_execution_preconditions | 已读AnchorLoop方法/实验：历史Executor frozenanchor同时给cross-referenceadvantage和Curriculum一致性proxy，VeRL sandbox+交替真实训练；curriculum时anchor与executor同参数，不能冒称版本改善或correctness。下一批须保留dualagent训练和reference snapshot，不以selfconsistency当gold。 |
| 2610.09761 PARC-Loc: Text-to-Point-Cloud Localization with Partial Assignment and Relational Consistency | explicit_unimplemented | PARCpartialassignment+pairwise spatial关系和neighborboundary扩充，KITTI360/CityLoc；真实text/pointcloud特征与coarse/fine权重待实现。 |
| 2610.09624 How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression | implementation_or_execution_preconditions | 已读toolcall causalvector分析/跨model评测：500minimalverbpairs300discover200heldout，hiddenstateintervention/transcoder，7families与native tau2 generalization。下一批须真实activations causalnecessity/sufficiency，不用文字heuristic决定toolcall或按test筛vector。 |
| 2610.09484 MIMESIS: Learning User Simulators as Training Environments for Interactive Agents | 完整实验不复现，审查DONE；原理由类型=implementation_or_execution_preconditions | Meta优先。已读MIMESIS user simulator及agent posttraining：user-side midtrain、ThoughtTrace SFT、多域GRPO后冻结simulator，再agent GRPO+CSD token重加权；private thought仅retrospective teacher训练上下文。9 unseen user评测和8环境，下一批须真实双训练链而非fixed simulator。 |
| 2610.09430 SkillCycle: Co-Evolving Agent Policies and Skill Banks | explicit_unimplemented | SkillCycle bank/router冻结policy学习后冻结policy修rules，rule/bank环境配对验证；ALFWorld/WebShop真实controller/distill待实现，no-skill部署不可带bank。 |
| 2610.09252 Adaptive Visual Token Reduction for Accelerated Image Understanding | explicit_unimplemented | ReFIT multi-scale instructionwindowreshape+tokenrefine，只训reductionmodules；四VQA FLOPs非wallclock提升，真实feature模块待实现。 |
| 2610.09159 SpecGuard: Proving a Task Is Broken Before the Agent Cheats | 完整实验不复现，审查DONE；原理由类型=integrity_or_acceptance_audit_preconditions | Google DeepMind优先。已读SpecGuard方法、实验和限制：SpecAgent独立测试生成Lean specification，TestAgent分离，semantic connector与kernel证书只有在忠实autoformalization下保证。349 mutated SWE任务和手工60规范审计，不是线上A/B；下一批须防规范弱化/伪冲突。 |
| 2610.09127 CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use | implementation_or_execution_preconditions | 已读CADFather method/experiment：VLM看target/intermediate renders，协调pretrainedCADgeneration/numericalrefinement，真CadQueryDSLexecute feedback与bestvalidarchive，无新training。下一批须真实3D几何/多tool策略，注意testsubsetablation不能参与parameter选择。 |
| 2610.09027 Visual Memory Attacks Can Persist Through The KV Cache | explicit_unimplemented | P-VMI whitebox图像攻击移除后仍经后续KV影响，cache-swap/summarycompaction控制；真KV交互/梯度攻击待实现，onlysoftprompt是diagnostic。 |