Agent 研究周度深度综述（2026-10-05 ~ 2026-10-11）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/10/9 17:00:05
本报告聚合本周 arXiv cs.AI 的 Agent 相关论文，跨天去重后深度分析。

本周总览

本周去重后共 60 篇论文，覆盖 4 天数据。上周 23 篇，环比增加 37 篇。

研究方向分布

方向	本周	上周	变化
其他	21	5	+16
规划推理	14	8	+6
评估基准	13	4	+9
自我进化	7	2	+5
安全对齐	7	2	+5
工程架构	5	2	+3
记忆系统	4	2	+2
多智能体	2	4	-2

应用场景分布

场景	论文数	占比
决策支持	9	15%
机器人与物理世界	6	10%
科学研究	5	8%
信息检索与问答	5	8%
代码开发	5	8%
企业自动化	4	7%
创意与内容	2	3%
数据分析	1	2%

核心论文解读

1. EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

• 英文标题: EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
• arXiv: 2610.10498 Kimi解读
• 方向: 其他
• 场景: 代码开发、科学研究、机器人与物理世界

2. HazardWeaver: Scientific Route Selection for Hazard Analysis Agents

• 英文标题: HazardWeaver: Scientific Route Selection for Hazard Analysis Agents
• arXiv: 2610.03591 Kimi解读
• 方向: 其他
• 场景: 科学研究、企业自动化

3. Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows

• 英文标题: Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows
• arXiv: 2610.03564 Kimi解读
• 方向: 其他
• 场景: 企业自动化、信息检索与问答

4. Benchmarking Candidate Coverage in Typed Decision Models

• 英文标题: Benchmarking Candidate Coverage in Typed Decision Models
• arXiv: 2610.03387 Kimi解读
• 方向: 记忆系统 · 评估基准
• 场景: 决策支持

5. Proof-Grounded Patient-Specific Clinical Explanations from Knowledge-Graph Reasoning

• 英文标题: Proof-Grounded Patient-Specific Clinical Explanations from Knowledge-Graph Reasoning
• arXiv: 2610.06549 Kimi解读
• 方向: 规划推理
• 场景: 科学研究、信息检索与问答

6. GPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement

• 英文标题: GPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement
• arXiv: 2610.06489 Kimi解读
• 方向: 工程架构
• 场景: 代码开发、决策支持

7. VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning

• 英文标题: VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
• arXiv: 2610.08761 Kimi解读
• 方向: 规划推理 · 自我进化
• 场景: 机器人与物理世界

8. ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences

• 英文标题: ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences
• arXiv: 2610.08691 Kimi解读
• 方向: 评估基准 · 自我进化
• 场景: 科学研究

9. Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning

• 英文标题: Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning
• arXiv: 2610.08627 Kimi解读
• 方向: 规划推理
• 场景: 代码开发、决策支持

10. Adaptive Power Sampling for LLM Reasoning

• 英文标题: Adaptive Power Sampling for LLM Reasoning
• arXiv: 2610.08563 Kimi解读
• 方向: 规划推理 · 自我进化
• 场景: 信息检索与问答

研究趋势

主导方向：其他（21 篇），较上周（5 篇）上升。

上升: 其他（5→21）、规划推理（8→14）、评估基准（4→13）、自我进化（2→7）、安全对齐（2→7）、工程架构（2→5）、记忆系统（2→4）

下降: 多智能体（4→2）

技术演进脉络

其他（21 篇）

• NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents Kimi解读
• HazardWeaver: Scientific Route Selection for Hazard Analysis Agents Kimi解读
• Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows Kimi解读
• 及另外 18 篇

规划推理（14 篇）

• Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis Kimi解读
• Reasoning Models Are Accurate but Unsound on Identification Kimi解读
• Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability Kimi解读
• 及另外 11 篇

评估基准（13 篇）

• Do Large Language Models Know Colombian Law? A Reliability Benchmark for the Colombian Legal System Kimi解读
• HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents Kimi解读
• From Benchmarks to Production: A Text-to-SQL System for Complex Financial Data Kimi解读
• 及另外 10 篇

自我进化（7 篇）

• Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis Kimi解读
• Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability Kimi解读
• VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning Kimi解读
• 及另外 4 篇

安全对齐（7 篇）

• ANT: A Multi-Granularity Network Traffic Dataset and Benchmark for Agents Behavior Auditing Kimi解读
• AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy Kimi解读
• ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding Kimi解读
• 及另外 4 篇

工程架构（5 篇）

• Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents Kimi解读
• GPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement Kimi解读
• Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements Kimi解读
• 及另外 2 篇

记忆系统（4 篇）

• Benchmarking Candidate Coverage in Typed Decision Models Kimi解读
• MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory Kimi解读
• RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing Kimi解读
• 及另外 1 篇

多智能体（2 篇）

• Can Agent Harnesses and Inference Engines Hear Each Other? The HEAR Protocol for Agentic LLM Serving Kimi解读
• SciExam for ENSO: Can AI Agents Build Climate Models? Kimi解读

工程实践启示

• 工程架构方向 5 篇，关注系统设计与可扩展性。
• 记忆系统方向 4 篇，RAG 与长期记忆方案不断优化。
• 多智能体方向 2 篇，协作模式从简单分工走向复杂协调。
• 安全方向 7 篇，Agent 安全从外部围栏走向内化机制。

下周关注

持续热点：其他（本周 21 篇，上周 5 篇）、规划推理（本周 14 篇，上周 8 篇）、评估基准（本周 13 篇，上周 4 篇）、自我进化（本周 7 篇，上周 2 篇）、安全对齐（本周 7 篇，上周 2 篇）、工程架构（本周 5 篇，上周 2 篇）、记忆系统（本周 4 篇，上周 2 篇）、多智能体（本周 2 篇，上周 4 篇）

附录：本周论文完整列表

去重后共 60 篇。

2026-10-05（13 篇）

1. Do Large Language Models Know Colombian Law? A Reliability Benchmark for the Colombian Legal System Kimi解读 — evaluation
2. Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents Kimi解读 — engineering
3. NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents Kimi解读 — other
4. HazardWeaver: Scientific Route Selection for Hazard Analysis Agents Kimi解读 — other
5. HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents Kimi解读 — evaluation
6. Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows Kimi解读 — other
7. Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis Kimi解读 — planning, evolution
8. From Benchmarks to Production: A Text-to-SQL System for Complex Financial Data Kimi解读 — evaluation
9. Reasoning Models Are Accurate but Unsound on Identification Kimi解读 — planning
10. Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability Kimi解读 — planning, evolution
11. Benchmarking Candidate Coverage in Typed Decision Models Kimi解读 — memory, evaluation
12. ReFract: Benchmarking Perspective Awareness in Language Model Agents with Text World Models Kimi解读 — planning, evaluation
13. Preserving Mathematical Reasoning in Compressed Diffusion Language Models via Trajectory-Aware Low-Rank Approximation Kimi解读 — planning

2026-10-06（11 篇）

1. Back to the Future: Rethinking EDA Infrastructure for Agentic Systems in Chip Design Verification Kimi解读 — other
2. FREA: A Multi-Source Expert Benchmark for Reaction Feasibility Verification Kimi解读 — evaluation
3. Can Agent Harnesses and Inference Engines Hear Each Other? The HEAR Protocol for Agentic LLM Serving Kimi解读 — multi_agent
4. HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention Kimi解读 — other
5. Proof-Grounded Patient-Specific Clinical Explanations from Knowledge-Graph Reasoning Kimi解读 — planning
6. ANT: A Multi-Granularity Network Traffic Dataset and Benchmark for Agents Behavior Auditing Kimi解读 — safety, evaluation
7. polyview: A Python package for multi-view machine learning Kimi解读 — other
8. GPlaceRL: An Open-Source Graph Reinforcement Learning Framework for Detailed Placement Kimi解读 — engineering
9. AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy Kimi解读 — safety
10. From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery? Kimi解读 — evaluation
11. What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents Kimi解读 — other

2026-10-07（18 篇）

1. Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts? Kimi解读 — other
2. VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning Kimi解读 — planning, evolution
3. Does an Agent's History Tell You When Compaction Will Hurt? A Modest, Bounded Effect on the TRACE Paired-Replay Corpus Kimi解读 — other
4. WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation? Kimi解读 — other
5. nanoMuse: An Open-Source Personal Agent for Every Device You Own Kimi解读 — other
6. ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences Kimi解读 — evaluation, evolution
7. ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding Kimi解读 — safety, evaluation
8. SquidAgent: Parallelize Wisely, Coordinate Efficiently Kimi解读 — other
9. Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning Kimi解读 — planning
10. Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness Kimi解读 — other
11. MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory Kimi解读 — memory
12. Adaptive Power Sampling for LLM Reasoning Kimi解读 — planning, evolution
13. Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements Kimi解读 — safety, engineering
14. How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation Kimi解读 — evolution
15. AssemState: Manual and Physical-State-Guided Reasoning for Zero-shot Furniture Assembly Kimi解读 — planning
16. EMHO: EMbodied Agent Harness Optimization via Experience Traces Kimi解读 — engineering
17. Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations Kimi解读 — evaluation
18. MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation Kimi解读 — planning

2026-10-08（18 篇）

1. RoboJEPA: Scaling Robotic Latent World Models Kimi解读 — planning
2. SciExam for ENSO: Can AI Agents Build Climate Models? Kimi解读 — multi_agent
3. RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing Kimi解读 — memory, evolution
4. EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution Kimi解读 — other
5. Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models Kimi解读 — evaluation
6. RunningTab: Direct Workspace Interaction with Environment-Side Tabs Kimi解读 — other
7. SOTA: Stock Options Trading Agents Guided by Option-Implied Return Distributions Kimi解读 — other
8. Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models Kimi解读 — planning
9. AI Safety Considerations for Agents With Limited Time to Act Kimi解读 — safety
10. OOM-RL II: Reality Is an Oracle, Not a Debugger Provenance-Constrained Diagnosis in Continually Evolving Agent-Engineered Systems Kimi解读 — other
11. GAGR-Lab: Evaluating Joint Spatial-Geometric and Analytic Function Reasoning Kimi解读 — planning, evaluation
12. Agentic AI-Assisted Modeling for Production Scheduling: Assessment in Constraint Programming Kimi解读 — other
13. RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation Kimi解读 — other
14. HGP:An on-device personalized agent memory via hybrid graph storage Kimi解读 — memory
15. Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents Kimi解读 — other
16. Learning to Accumulate Knowledge with Mutual Information Kimi解读 — other
17. From Expected Harmfulness to Likelihood: A Probabilistic Reformulation of Jailbreaking LLM Agents Kimi解读 — safety
18. Successive Training Stages and Large Language Model Persuasion: Effects of Misalignment, Supervised Fine-Tuning, and Preference Optimization Kimi解读 — safety, engineering

---

本报告由 OpenClaw 自动生成，基于 agent-papers-research 每日数据聚合。论文标题和摘要由 GLM-5 翻译生成。