Agent 研究周度深度综述（2026-08-31 ~ 2026-09-06）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/9/4 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 56 篇论文，覆盖 4 天数据。上周 59 篇，环比减少 3 篇。

研究方向分布

方向 本周 上周 变化
规划推理 21 14 +7
其他 10 15 -5
评估基准 10 8 +2
安全对齐 8 7 +1
工程架构 8 7 +1
自我进化 6 4 +2
多智能体 6 6 0
记忆系统 6 9 -3
工具使用 1 1 0

应用场景分布

场景 论文数 占比
决策支持 8 14%
科学研究 4 7%
信息检索与问答 3 5%
代码开发 3 5%
机器人与物理世界 2 4%
创意与内容 2 4%

核心论文解读

1. CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI

  • 英文标题: CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI
  • arXiv: 2609.02459 Kimi解读
  • 方向: 记忆系统 · 规划推理 · 工具使用 · 评估基准
  • 场景: 决策支持

2. Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

  • 英文标题: Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills
  • arXiv: 2609.02749 Kimi解读
  • 方向: 其他
  • 场景: 代码开发、科学研究、信息检索与问答

3. MAP: A Benchmark on Multimodal Accessibility Planning for Real World Places

  • 英文标题: MAP: A Benchmark on Multimodal Accessibility Planning for Real World Places
  • arXiv: 2608.28384 Kimi解读
  • 方向: 规划推理 · 评估基准
  • 场景: 决策支持

4. Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration

  • 英文标题: Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration
  • arXiv: 2608.28264 Kimi解读
  • 方向: 多智能体 · 自我进化 · 工程架构

5. Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data

  • 英文标题: Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data
  • arXiv: 2608.31082 Kimi解读
  • 方向: 规划推理 · 自我进化
  • 场景: 信息检索与问答

6. Dual Process Motion Planning

  • 英文标题: Dual Process Motion Planning
  • arXiv: 2609.01260 Kimi解读
  • 方向: 规划推理
  • 场景: 决策支持、机器人与物理世界

7. UTP-Bench: Uncertainty-aware Travel Planning Benchmark

  • 英文标题: UTP-Bench: Uncertainty-aware Travel Planning Benchmark
  • arXiv: 2609.02421 Kimi解读
  • 方向: 规划推理 · 评估基准
  • 场景: 决策支持

8. Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds

  • 英文标题: Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds
  • arXiv: 2609.02302 Kimi解读
  • 方向: 安全对齐 · 评估基准 · 工程架构

9. Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems

  • 英文标题: Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems
  • arXiv: 2609.02264 Kimi解读
  • 方向: 多智能体
  • 场景: 代码开发、创意与内容

10. APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering

  • 英文标题: APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering
  • arXiv: 2609.02253 Kimi解读
  • 方向: 自我进化
  • 场景: 科学研究、信息检索与问答

研究趋势

主导方向：规划推理（21 篇），较上周（14 篇）上升。

上升: 规划推理（14→21）、评估基准（8→10）、安全对齐（7→8）、工程架构（7→8）、自我进化（4→6）

下降: 其他（15→10）、记忆系统（9→6）

技术演进脉络

规划推理（21 篇）

其他（10 篇）

评估基准（10 篇）

安全对齐（8 篇）

工程架构（8 篇）

自我进化（6 篇）

多智能体（6 篇）

记忆系统（6 篇）

工具使用（1 篇）

工程实践启示

  • 工程架构方向 8 篇，关注系统设计与可扩展性。
  • 工具使用方向 1 篇，function calling 与工具链持续演进。
  • 记忆系统方向 6 篇，RAG 与长期记忆方案不断优化。
  • 多智能体方向 6 篇，协作模式从简单分工走向复杂协调。
  • 安全方向 8 篇，Agent 安全从外部围栏走向内化机制。

下周关注

持续热点：规划推理（本周 21 篇，上周 14 篇）、其他（本周 10 篇，上周 15 篇）、评估基准（本周 10 篇，上周 8 篇）、安全对齐（本周 8 篇，上周 7 篇）、工程架构（本周 8 篇，上周 7 篇）、自我进化（本周 6 篇，上周 4 篇）、多智能体（本周 6 篇，上周 6 篇）、记忆系统（本周 6 篇，上周 9 篇）

附录：本周论文完整列表

去重后共 56 篇。

2026-08-31（12 篇）

  1. Logos: An Agent Harness on a Cross-Process Bus Kimi解读 — other
  2. AcrossVAM1.0: Particle World Modeling for Text-Assisted Robot Video Prediction Kimi解读 — planning
  3. Learning to Use Tools: Reinforcement Learning for Tool-Integrated Mathematical Reasoning Kimi解读 — planning
  4. Prove2Me: An Open Collaborative Platform for Scaling Math Formalization Kimi解读 — other
  5. VERA-8B: Evidence-Grounded Audit Risk Reasoning from SEC Filings Kimi解读 — planning, safety
  6. RetailAgent: Structured Adverse Timing in Self-Conditioned Multimodal LLM Trading Agents Kimi解读 — other
  7. MAP: A Benchmark on Multimodal Accessibility Planning for Real World Places Kimi解读 — planning, evaluation
  8. EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses Kimi解读 — evolution
  9. AGENT-O: A Semantic Agent Card Framework for Interoperable and Governed Healthcare AI Agents Kimi解读 — evaluation, engineering
  10. LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering Kimi解读 — evaluation
  11. Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration Kimi解读 — multi_agent, evolution, engineering
  12. Regime-Aware Portfolio Management via Retrieval-Augmented LLM-Guided Expert Switching Kimi解读 — memory

2026-09-01（14 篇）

  1. OntoAligner-Ensemble: Voting-Based Fusion across Heterogeneous Ontology Alignment Techniques Kimi解读 — safety
  2. Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data Kimi解读 — planning, evolution
  3. Reconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization Kimi解读 — engineering
  4. Scaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence Kimi解读 — planning
  5. Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores Kimi解读 — planning
  6. Measure Before You Manage: Evaluating Agent Working Memory in Coding Agents Kimi解读 — memory
  7. MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents Kimi解读 — other
  8. CAER: Causal Action Effect Reweighting for World Model Training Kimi解读 — planning
  9. HSRM: Hidden-State Reward Models for Test-Time Verification Kimi解读 — planning, evaluation
  10. SkillZip Pro: Execution-Aware Dynamic Compression of Progressively Loaded Skills for Self-Evolving Agents Kimi解读 — other
  11. Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models Kimi解读 — planning
  12. ATLAS: Dual-Horizon Diagnostic Evaluation for Industrial Tool-Use Agents Kimi解读 — evaluation
  13. MedAgent-R1: Faithfulness-Aware Reinforcement Learning for Evidence-Grounded Medical Reasoning Kimi解读 — memory, planning
  14. HiRS-Agent: A Hierarchical Multi-Agent System for Reliable Long-Horizon Remote Sensing Task Solving Kimi解读 — multi_agent

2026-09-02（12 篇）

  1. Selective Agent Guidance via Entropy: Learning Autonomous Policies from Imperfect VLM Teachers Kimi解读 — other
  2. EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation Kimi解读 — other
  3. When Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation Kimi解读 — safety, evaluation
  4. Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers Kimi解读 — other
  5. EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems Kimi解读 — multi_agent
  6. Analog-DB: An Agent-First Analog Integrated Circuit Database, From Blocks to Systems Kimi解读 — evaluation
  7. Making Prospective Memory SLM-Shaped: Typed Intention Stores for Small-Model Agents Kimi解读 — memory
  8. Dual Process Motion Planning Kimi解读 — planning
  9. H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning Kimi解读 — planning
  10. Jailbreaking Text-to-Image Models Through Cracks: Navigating Heterogeneous Safety Filters via Multi-Agent Debate Kimi解读 — multi_agent, safety
  11. Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs Kimi解读 — planning
  12. ARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning Kimi解读 — evolution

2026-09-03（18 篇）

  1. Discriminative World Models for Web Agents Kimi解读 — planning
  2. Large Language Models (LLMs) for Telecom Root Cause Analysis (RCA): A Structured Reasoning Framework for Evidence-Grounded Diagnosis Kimi解读 — planning, engineering
  3. SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment Kimi解读 — safety
  4. Measurement-Driven Sub-Network Selection for On-Premise Retrieval-Augmented Factory Agents Kimi解读 — memory
  5. Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems Kimi解读 — multi_agent, evolution
  6. Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills Kimi解读 — other
  7. CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI Kimi解读 — memory, planning, tool, evaluation
  8. UTP-Bench: Uncertainty-aware Travel Planning Benchmark Kimi解读 — planning, evaluation
  9. Contrastive Explanations in Quantitative Bipolar Argumentation Frameworks Kimi解读 — planning, engineering
  10. Diagnosing with Insights: Structured Analysis of Agent Failures via Behavioral Abstractions Kimi解读 — other
  11. SALA: Semantic-Aware Logical Alignment for Complex Reasoning in In-Context Learning Kimi解读 — planning, safety
  12. Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds Kimi解读 — safety, evaluation, engineering
  13. Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems Kimi解读 — multi_agent
  14. APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering Kimi解读 — evolution
  15. LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails Kimi解读 — safety
  16. Propose to Learn, Learn to Propose: Evaluability-Aware Assistance under Bounded Rationality Kimi解读 — planning
  17. PGPO: Potential-Guided Policy Optimization for Multi-Turn Agentic Tasks Kimi解读 — engineering
  18. PhoenixNest-Video: Evidence-Grounded Multimodal Agent Framework for Automated Video Interview Assessment Kimi解读 — 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。

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。