Agent 研究周度深度综述（2026-10-26 ~ 2026-11-01）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/10/30 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 32 篇论文，覆盖 3 天数据。上周 74 篇，环比减少 42 篇。

研究方向分布

方向 本周 上周 变化
其他 12 21 -9
记忆系统 7 13 -6
规划推理 6 26 -20
评估基准 4 10 -6
多智能体 2 3 -1
安全对齐 2 1 +1
工程架构 2 9 -7
自我进化 1 3 -2

应用场景分布

场景 论文数 占比
信息检索与问答 5 16%
决策支持 4 13%
代码开发 4 13%
企业自动化 3 9%
科学研究 2 6%
创意与内容 1 3%

核心论文解读

1. Reimagining Enterprise RAG for Multi-Agent Systems: A Knowledge Layer That Organizes Before It Retrieves

  • 英文标题: Reimagining Enterprise RAG for Multi-Agent Systems: A Knowledge Layer That Organizes Before It Retrieves
  • arXiv: 2610.32851 Kimi解读
  • 方向: 记忆系统 · 多智能体
  • 场景: 企业自动化、信息检索与问答

2. What Transfers Across Protein Pockets? Experience Reuse in Molecular Optimization for Agentic Structure-Based Drug Design

  • 英文标题: What Transfers Across Protein Pockets? Experience Reuse in Molecular Optimization for Agentic Structure-Based Drug Design
  • arXiv: 2610.32620 Kimi解读
  • 方向: 工程架构
  • 场景: 创意与内容、决策支持

3. POSEIDON: Scaling Training Environments for Scientific Agents from Code Repositories

  • 英文标题: POSEIDON: Scaling Training Environments for Scientific Agents from Code Repositories
  • arXiv: 2610.32448 Kimi解读
  • 方向: 其他
  • 场景: 代码开发、科学研究

4. Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation

  • 英文标题: Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation
  • arXiv: 2610.37242 Kimi解读
  • 方向: 其他
  • 场景: 企业自动化、信息检索与问答

5. LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents

  • 英文标题: LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents
  • arXiv: 2610.37166 Kimi解读
  • 方向: 评估基准
  • 场景: 代码开发、信息检索与问答

6. Prospector: learning a compact graph traversal policy for scientific reasoning

  • 英文标题: Prospector: learning a compact graph traversal policy for scientific reasoning
  • arXiv: 2610.37121 Kimi解读
  • 方向: 规划推理
  • 场景: 科学研究、信息检索与问答

7. HarnessMM: Recursive Self-Improvement for Harness Token Efficiency

  • 英文标题: HarnessMM: Recursive Self-Improvement for Harness Token Efficiency
  • arXiv: 2610.32813 Kimi解读
  • 方向: 自我进化
  • 场景: 决策支持

8. Agentic Hazard Mapper: Emulating Expert Map-Making with AI Agents for Risk Communication

  • 英文标题: Agentic Hazard Mapper: Emulating Expert Map-Making with AI Agents for Risk Communication
  • arXiv: 2610.32738 Kimi解读
  • 方向: 多智能体 · 安全对齐

9. LGMT: A Tool for Testing the Logical Reasoning Reliability of LLMs

  • 英文标题: LGMT: A Tool for Testing the Logical Reasoning Reliability of LLMs
  • arXiv: 2610.32627 Kimi解读
  • 方向: 规划推理 · 评估基准

10. When Agents Fail: Model or Harness–Who's to Blame?

  • 英文标题: When Agents Fail: Model or Harness–Who's to Blame?
  • arXiv: 2610.32483 Kimi解读
  • 方向: 其他
  • 场景: 代码开发

研究趋势

主导方向：其他（12 篇），较上周（21 篇）下降。

上升: 安全对齐（1→2）

下降: 其他（21→12）、规划推理（26→6）、工程架构（9→2）、记忆系统（13→7）、自我进化（3→1）、工具使用（2→0）、评估基准（10→4）、多智能体（3→2）

技术演进脉络

其他（12 篇）

记忆系统（7 篇）

规划推理（6 篇）

评估基准（4 篇）

多智能体（2 篇）

安全对齐（2 篇）

工程架构（2 篇）

自我进化（1 篇）

工程实践启示

  • 工程架构方向 2 篇，关注系统设计与可扩展性。
  • 记忆系统方向 7 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 2 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 12 篇，上周 21 篇）、记忆系统（本周 7 篇，上周 13 篇）、规划推理（本周 6 篇，上周 26 篇）、评估基准（本周 4 篇，上周 10 篇）、多智能体（本周 2 篇，上周 3 篇）、工程架构（本周 2 篇，上周 9 篇）

附录：本周论文完整列表

去重后共 32 篇。

2026-10-26（13 篇）

  1. Closing the Modality Gap via Simplex-Constrained Representations Kimi解读 — memory
  2. Do "True" and "False" Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations Kimi解读 — other
  3. Reimagining Enterprise RAG for Multi-Agent Systems: A Knowledge Layer That Organizes Before It Retrieves Kimi解读 — memory, multi_agent
  4. HarnessMM: Recursive Self-Improvement for Harness Token Efficiency Kimi解读 — evolution
  5. ActionRAG: Evidence-Refining Graph Diffusion for Long-Term Memory in LLM Agents Kimi解读 — memory
  6. Agentic Hazard Mapper: Emulating Expert Map-Making with AI Agents for Risk Communication Kimi解读 — multi_agent, safety
  7. LGMT: A Tool for Testing the Logical Reasoning Reliability of LLMs Kimi解读 — planning, evaluation
  8. What Transfers Across Protein Pockets? Experience Reuse in Molecular Optimization for Agentic Structure-Based Drug Design Kimi解读 — engineering
  9. IGG: A Benchmark for Interactive GUI Grounding under Visibility Constraints Kimi解读 — evaluation
  10. When Agents Fail: Model or Harness–Who's to Blame? Kimi解读 — other
  11. CauProphet: Teaching Agents to Develop Causal Cognition like Humans Kimi解读 — other
  12. Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings Kimi解读 — memory, planning
  13. POSEIDON: Scaling Training Environments for Scientific Agents from Code Repositories Kimi解读 — other

2026-10-28（19 篇）

  1. KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry Kimi解读 — memory
  2. TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information Kimi解读 — other
  3. From Interaction to Capability: Task Advancement as State and Supervision for GUI Agents Kimi解读 — other
  4. Concordat: Typed Sessions for LLM Agents Kimi解读 — other
  5. M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents Kimi解读 — other
  6. Constraining the Commit Boundary: A DIO/ODI-Informed Authorization Guardrail for Agentic Entity Resolution Kimi解读 — safety
  7. Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their Answers Kimi解读 — other
  8. Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model Kimi解读 — planning
  9. Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories Kimi解读 — planning
  10. Aligning Institutional Power and Operational Ability Kimi解读 — other
  11. ReForest: Interpretability Meets Reasoning in Isolation Forest Kimi解读 — planning
  12. RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal Transport Kimi解读 — memory
  13. TCM-DiagCommit: A TCM-SD-Derived Benchmark for Evidence-Conditioned Syndrome Selection and Commitment Kimi解读 — evaluation
  14. Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation Kimi解读 — other
  15. GraphRAGProv: GraphRAG Meets How-Provenance Kimi解读 — memory
  16. Does Your Generalist Skill Optimizer Really Incentivize Agentic Tasks? Rethinking Skill Optimization via Bipartite Attribution Kimi解读 — engineering
  17. LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents Kimi解读 — evaluation
  18. WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training Kimi解读 — other
  19. Prospector: learning a compact graph traversal policy for scientific reasoning Kimi解读 — planning

2026-10-29（19 篇）

  1. KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry Kimi解读 — memory
  2. TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information Kimi解读 — other
  3. From Interaction to Capability: Task Advancement as State and Supervision for GUI Agents Kimi解读 — other
  4. Concordat: Typed Sessions for LLM Agents Kimi解读 — other
  5. M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents Kimi解读 — other
  6. Constraining the Commit Boundary: A DIO/ODI-Informed Authorization Guardrail for Agentic Entity Resolution Kimi解读 — safety
  7. Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their Answers Kimi解读 — other
  8. Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model Kimi解读 — planning
  9. Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories Kimi解读 — planning
  10. Aligning Institutional Power and Operational Ability Kimi解读 — other
  11. ReForest: Interpretability Meets Reasoning in Isolation Forest Kimi解读 — planning
  12. RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal Transport Kimi解读 — memory
  13. TCM-DiagCommit: A TCM-SD-Derived Benchmark for Evidence-Conditioned Syndrome Selection and Commitment Kimi解读 — evaluation
  14. Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation Kimi解读 — other
  15. GraphRAGProv: GraphRAG Meets How-Provenance Kimi解读 — memory
  16. Does Your Generalist Skill Optimizer Really Incentivize Agentic Tasks? Rethinking Skill Optimization via Bipartite Attribution Kimi解读 — engineering
  17. LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents Kimi解读 — evaluation
  18. WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training Kimi解读 — other
  19. Prospector: learning a compact graph traversal policy for scientific reasoning Kimi解读 — planning

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。

相关文章推荐

tag:

    缺失模块。
    1、请确保node版本大于6.2
    2、在博客根目录（注意不是yilia根目录）执行以下命令：
    npm i hexo-generator-json-content --save

    3、在根目录_config.yml里添加配置： 

      jsonContent:
    meta: false
    pages: false
    posts:
      title: true
      date: true
      path: true
      text: false
      raw: false
      content: false
      slug: false
      updated: false
      comments: false
      link: false
      permalink: false
      excerpt: false
      categories: false
      tags: true

程序员，关于技术我所知甚少，唯有探索，方得真知。