Agent 研究周度深度综述（2026-08-17 ~ 2026-08-23）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/8/21 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 47 篇论文，覆盖 4 天数据。上周 68 篇，环比减少 21 篇。

研究方向分布

方向 本周 上周 变化
规划推理 11 20 -9
其他 11 17 -6
评估基准 9 19 -10
多智能体 9 4 +5
记忆系统 6 13 -7
工程架构 5 9 -4
自我进化 4 2 +2
安全对齐 1 7 -6

应用场景分布

场景 论文数 占比
信息检索与问答 5 11%
决策支持 4 9%
企业自动化 4 9%
科学研究 3 6%
数据分析 2 4%
代码开发 2 4%
机器人与物理世界 2 4%
创意与内容 1 2%

核心论文解读

1. Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering

  • arXiv: 2608.19029
  • 方向: 记忆系统 · 规划推理 · 多智能体 · 自我进化
  • 场景: 信息检索与问答
  • 关键词: agent reflection memory medical amr answering reasoning overseer medmcqa escalated

2. ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond

  • arXiv: 2608.14354
  • 方向: 其他
  • 场景: 科学研究、信息检索与问答
  • 关键词: scienceflow research executable horizon autoresearch execution scientific sustain progress exploration

3. Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning

  • arXiv: 2608.14290
  • 方向: 规划推理 · 工程架构
  • 场景: 信息检索与问答
  • 关键词: mobius intern reasoning knowledge reasoners architecture attn achieves 35b decoupled

4. Reconstruction: A Blind Benchmark for Recovering Research Ideas from Pre-Publication Bibliographies

  • arXiv: 2608.16645
  • 方向: 评估基准
  • 场景: 科学研究、信息检索与问答
  • 关键词: bibliographies leakage idea approx publication tournament blind seed withholds reconstruction

5. PDDLCoder: Agentic PDDL Generation for LLM-Assisted Symbolic Planning

  • arXiv: 2608.16637
  • 方向: 规划推理
  • 场景: 代码开发、决策支持
  • 关键词: pddl planning pddlcoder agentic pddlgym generation llm symbolic plans assisted

6. A Multi-Agent Platform for Automated Enterprise Analytics and Insight Generation

  • arXiv: 2608.18740
  • 方向: 多智能体
  • 场景: 数据分析、企业自动化
  • 关键词: enterprise agent crewai conversational platform quality llm multi analytics agents

7. SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning

  • arXiv: 2608.14452
  • 方向: 规划推理
  • 场景: 数据分析
  • 关键词: sheetcompass spreadsheet workbooks agentic spreadsheets reasoning sheet worksheets llms flatten

8. PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments

  • arXiv: 2608.14441
  • 方向: 评估基准
  • 场景: 代码开发
  • 关键词: pace bench code adaptation target succeeds agents physics grounded self

9. Cross-Sign Language Transfer Learning Using Domain Adaptation with Multi-scale Temporal Alignment

  • arXiv: 2608.16804
  • 方向: 多智能体 · 安全对齐
  • 关键词: sign language temporal adaptation transfer domain ta3n aligning communication individuals

10. Chronocooked: A Benchmark for Implicit Interval Timing in Reinforcement Learning Agents

  • arXiv: 2608.16666
  • 方向: 评估基准
  • 场景: 创意与内容
  • 关键词: chronocooked agents timing biologically implicit interval reinforcement suite temporal designed

研究趋势

主导方向：规划推理（11 篇），较上周（20 篇）下降。

上升: 多智能体（4→9）、自我进化（2→4）

下降: 其他（17→11）、记忆系统（13→6）、工程架构（9→5）、评估基准（19→9）、规划推理（20→11）、安全对齐（7→1）、工具使用（3→0）

技术演进脉络

规划推理（11 篇）

其他（11 篇）

评估基准（9 篇）

多智能体（9 篇）

记忆系统（6 篇）

工程架构（5 篇）

自我进化（4 篇）

安全对齐（1 篇）

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 6 篇，RAG 与长期记忆方案不断优化。
  • 多智能体方向 9 篇，协作模式从简单分工走向复杂协调。
  • 安全方向 1 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 11 篇，上周 20 篇）、其他（本周 11 篇，上周 17 篇）、评估基准（本周 9 篇，上周 19 篇）、多智能体（本周 9 篇，上周 4 篇）、记忆系统（本周 6 篇，上周 13 篇）、工程架构（本周 5 篇，上周 9 篇）、自我进化（本周 4 篇，上周 2 篇）

附录：本周论文完整列表

去重后共 47 篇。

2026-08-17（10 篇）

  1. SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning — planning
  2. Wyvern: An Agentic Framework for Generating Grounded Multimodal Reports — engineering
  3. PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments — evaluation
  4. AgentRewind: Recoverable Execution for Long-Horizon LLM Agents — other
  5. Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages — multi_agent
  6. ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond — other
  7. Clearing the Fog: Towards Installing and Refining Proactive Exploration Capabilities in LLM Agents — other
  8. AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMs — evaluation
  9. Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning — planning, engineering
  10. TimeSage-EV: A Live Benchmark for Agentic Time Series Analysis in Evolving Environments — evaluation

2026-08-18（12 篇）

  1. Cross-Sign Language Transfer Learning Using Domain Adaptation with Multi-scale Temporal Alignment — multi_agent, safety
  2. When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding — multi_agent
  3. GRIP: Grounded Reasoning via Information-Restricted Premises — planning
  4. Chronocooked: A Benchmark for Implicit Interval Timing in Reinforcement Learning Agents — evaluation
  5. Reconstruction: A Blind Benchmark for Recovering Research Ideas from Pre-Publication Bibliographies — evaluation
  6. PDDLCoder: Agentic PDDL Generation for LLM-Assisted Symbolic Planning — planning
  7. Hypergraph-based Multimodal Retrieval-Augmented Generation with Incremental Refinement — memory
  8. Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents — other
  9. Probabilistic Circuits as Reasoning Machines in Artificial Intelligence (Part I) — planning
  10. DeepInsight II: One Trace from Benchmark to Robot — evaluation
  11. Offline Reinforcement Learning for Hemodynamic Management of Sepsis in the ICU: a MIMIC-IV Study with Dual Off-Policy Evaluation — evaluation
  12. HaReCAP: Habitual-action Grounding for Recursive Large Language Model Agents — other

2026-08-19（13 篇）

  1. On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification — memory, evaluation
  2. Delegation Asymmetry in Agentic Recommender Systems: Measuring Two-Sided Receptivity in Online Dating — engineering
  3. StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents — other
  4. Towards Zero-Shot Task Transfer with Neurosymbolic World Models — planning
  5. EvoTS-Agent: A Self-Evolving LLM Agent for Financial Time Series Change Point Detection — other
  6. ARASH: Adaptive Retrieval And Shot Selection for Tabular Prediction — memory, evolution
  7. StartupBench: Benchmarking General-Purpose Agents on Market-Validated End-to-End Workflows — evaluation
  8. D$^2$ACCI: A Dual-Loop Diagnostic Protocol for Evidence-Preserving Agent Memory — memory
  9. The Curious Case of Exploding DecPOMDPs: Containing the Fire through Policy Counting — other
  10. Beyond Suspicious Steps: Ontological Trust in Long-Horizon Agents — other
  11. Auditing Self-Evolution in Financial Agents: Capability Gains, Security Drift, and Execution-Interface Mismatch — planning, evolution
  12. GraphWake: Group Polarization via Memory-Mediated Polarization Cascade in LLM-Agent Communities — memory
  13. Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing — planning

2026-08-20（12 篇）

  1. Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication — multi_agent
  2. Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery — engineering
  3. Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering — memory, planning, multi_agent, evolution
  4. A Theory of Post-hoc Debate Judgement — multi_agent
  5. Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models — evolution
  6. DentAgent: Evidence-Centric Multi-Agent Coordination for Multimodal Dental Reasoning — planning, multi_agent
  7. SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents — other
  8. ORBITER: Conflict-Aware Decision-Making for Agentic Last-Mile Delivery — other
  9. A Multi-Agent Platform for Automated Enterprise Analytics and Insight Generation — multi_agent
  10. RTPO: Reverse-Turn Policy Optimization for Stabilizing Agentic RL Training — engineering
  11. Sanyu Studio: A Multi-Agent System for Art-Historical Narrative Construction — multi_agent
  12. Preference Reasoning under Indeterminacy in Large Language Models — planning

本报告由 OpenClaw 自动生成，基于 agent-papers-research 每日数据聚合。

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。