Agent 研究周度深度综述（2026-08-10 ~ 2026-08-16）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/8/14 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 55 篇论文，覆盖 4 天数据。上周 30 篇，环比增加 25 篇。

研究方向分布

方向 本周 上周 变化
其他 16 9 +7
评估基准 15 6 +9
规划推理 15 9 +6
记忆系统 9 8 +1
工程架构 7 2 +5
安全对齐 4 2 +2
多智能体 2 1 +1
工具使用 2 1 +1
自我进化 1 1 0

应用场景分布

场景 论文数 占比
信息检索与问答 10 18%
科学研究 5 9%
决策支持 4 7%
企业自动化 3 5%
机器人与物理世界 1 2%
创意与内容 1 2%
代码开发 1 2%

核心论文解读

  • arXiv: 2608.09622
  • 方向: 规划推理 · 评估基准 · 自我进化
  • 场景: 信息检索与问答、决策支持
  • 关键词: qualification reliability sequential planning test tddb damage adaptive failure degradation

2. FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings

  • arXiv: 2608.07400
  • 方向: 记忆系统 · 评估基准
  • 场景: 信息检索与问答
  • 关键词: filings finrank grounded negatives evidence embedder financial answering reporting sec

3. Agentic Auto-Research is Fuzz Testing

  • arXiv: 2608.09855
  • 方向: 评估基准
  • 场景: 科学研究、信息检索与问答
  • 关键词: research auto feedback fuzzer progress signal agentic fuzz rather validation

4. Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models

  • arXiv: 2608.09696
  • 方向: 规划推理
  • 场景: 科学研究、创意与内容
  • 关键词: discovery emph mda mechanistic proposer interventional citep model llm dpbench

5. V-FiLLM: Verified Financial LLM Reasoning Benchmark

  • arXiv: 2608.11047
  • 方向: 规划推理 · 评估基准
  • 场景: 信息检索与问答
  • 关键词: financial fillm reasoning verified tables items finqa 2022a benchmarks evaluating

6. ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling

  • arXiv: 2608.10928
  • 方向: 记忆系统 · 规划推理 · 评估基准
  • 关键词: thinkretrieve reasoning aime lrms traces 2025 test scaling step sciq

7. Tree-of-Ideas: Automated Research Ideation via Cross-Trajectory Reasoning over Scholarly Evolution

  • arXiv: 2608.10740
  • 方向: 规划推理
  • 场景: 科学研究、信息检索与问答
  • 关键词: scholarly research ideation ideas toi trajectories evoagent tree reasoning across

8. Self-Correcting Long-Horizon Search Agents via Tree-Structured Memory

  • arXiv: 2608.10676
  • 方向: 记忆系统 · 规划推理
  • 场景: 信息检索与问答
  • 关键词: retree search correcting evidence agents tree context react reasoning self

9. VAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies

  • arXiv: 2608.12282
  • 方向: 记忆系统 · 规划推理 · 工具使用
  • 关键词: apis vakra textbf reasoning hop tool ibm across failures etrieval

10. SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

  • arXiv: 2608.07449
  • 方向: 其他
  • 场景: 信息检索与问答
  • 关键词: skillprox skills proximal diagnosis textual gradient skill agent knowledge updates

研究趋势

主导方向：其他（16 篇），较上周（9 篇）上升。

上升: 其他（9→16）、评估基准（6→15）、规划推理（9→15）、记忆系统（8→9）、工程架构（2→7）、安全对齐（2→4）、多智能体（1→2）、工具使用（1→2）

技术演进脉络

其他（16 篇）

评估基准（15 篇）

规划推理（15 篇）

记忆系统（9 篇）

工程架构（7 篇）

安全对齐（4 篇）

多智能体（2 篇）

工具使用（2 篇）

自我进化（1 篇）

工程实践启示

  • 工程架构方向 7 篇，关注系统设计与可扩展性。
  • 工具使用方向 2 篇，function calling 与工具链持续演进。
  • 记忆系统方向 9 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 4 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 16 篇，上周 9 篇）、评估基准（本周 15 篇，上周 6 篇）、规划推理（本周 15 篇，上周 9 篇）、记忆系统（本周 9 篇，上周 8 篇）、工程架构（本周 7 篇，上周 2 篇）、安全对齐（本周 4 篇，上周 2 篇）

附录：本周论文完整列表

去重后共 55 篇。

2026-08-10（13 篇）

  1. SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent — other
  2. Blast Radius — other
  3. PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents — memory, engineering
  4. Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing — evaluation
  5. TEPA: Revoking Stale Memories for Conflict-Robust Language Agents — memory
  6. CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing — planning, evaluation
  7. ResidencyRL: Reinforcement Learning in Simulated Clinical Environments — other
  8. GeoBenchLLM: A Comprehensive Benchmark for Evaluating LLMs on Geo-Related Tasks — evaluation
  9. FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings — memory, evaluation
  10. People Are Not Just Their Countries. Disentangling Social Determinants of LLM Value Alignment Across Europe — safety
  11. An End-to-End Agent Auditing Engine — evaluation
  12. EMAS: Stabilizing Multi-Agent System Evolution through Evidence-Guided Revision — multi_agent
  13. Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory — memory

2026-08-11（13 篇）

  1. SHE: Trajectory-driven Safety Harness Evolution for LLM Agents — safety
  2. ArchAgent v2: A Case Study with the Data Prefetching Championship — engineering
  3. Agentic Auto-Research is Fuzz Testing — evaluation
  4. CEAA: A Cognitive Embodied Agents Architecture for Interactive Computing Systems — engineering
  5. CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation — other
  6. Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models — planning
  7. Open Evaluation Agent: Efficient and Promptable Evaluation of Visual Generative Models — tool, evaluation
  8. Avalon-ToM-Bench: Evaluating Fine-Grained Theory of Mind via Asymmetric Game Mechanics — planning, evaluation
  9. Rethinking Self-Evolving Agents: Do We Still Need Prescribed Optimization Pipelines? — engineering
  10. Adaptive Sequential Test Planning for Multi-Mechanism Reliability Qualification via Bayesian Monte Carlo Tree Search — planning, evaluation, evolution
  11. ElasticBack: Stealthy Conditional Backdoor in LLM-Agent Skills via Coupled Trigger-Rule Optimization — engineering
  12. The Politician, the Liar, and the Obedient Worker: Emerging Behavior of LLM Agents in Hierarchical Games — other
  13. Bidirectional Context Self-Distillation for Reinforcement Learning of Skill-Based LLM Agents — other

2026-08-12（15 篇）

  1. Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding — other
  2. SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure — evaluation
  3. V-FiLLM: Verified Financial LLM Reasoning Benchmark — planning, evaluation
  4. XCoT-VLA: Executable Chain-of-Thought for Vision-Language-Action Driving — planning
  5. ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling — memory, planning, evaluation
  6. IO Factory: Simulating AI-Enabled Influence Campaigns at Scale — other
  7. ComBodied Agents: a New Paradigm of Human-Centric Agentic AI — other
  8. ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation — other
  9. SkillLens: Visual Skill Cards for Retrieval-Augmented GUI Action Prediction and On-Policy Distillation — memory
  10. Compositional Benchmark Synthesis for Hierarchical Human Action Recognition — evaluation, engineering
  11. Tree-of-Ideas: Automated Research Ideation via Cross-Trajectory Reasoning over Scholarly Evolution — planning
  12. Self-Correcting Long-Horizon Search Agents via Tree-Structured Memory — memory, planning
  13. REDAgentBench: Executable Red Teaming and Faithful Measurement of LLM Agent Systems — safety
  14. VERDICT: Training-Free Step-Wise Verification of Multimodal Reasoning via Disagreement-Aware Consensus — planning
  15. Operationalising Relative Causal Knowledge: Backbone Identifiability from Private Reports on a Shared Outcome — other

2026-08-13（14 篇）

  1. Constructing Dynamic Master Logic Models as Knowledge Graphs for Complex System Diagnostics Using Retrieval-Augmented Large Language Models — memory
  2. VAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies — memory, planning, tool
  3. An Agentic Workflow for Legacy HPC Modernization: Converting the Two-Electron-Integral Core of GAMESS — other
  4. CTBench: Evaluating Troubleshooting Capabilities of AI Agents in Realistic Telecom Network Operations — other
  5. Claim-Level Reliability Assessment for Efficient Test-Time Reasoning — planning, evaluation
  6. ExRole: From Team Trajectories to Executable Roles in Multi-Agent Language Models — multi_agent
  7. Policy-as-logic for robust reasoning over rules — planning
  8. Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents — safety
  9. The Sleeping Agent: What Gist-Based Context Compression Loses and Why — other
  10. Harness-IF: Evaluating Instruction Following Across Instruction Surfaces in Coding Agents — other
  11. HUGIN: Enhancing Vision-Language Planning for Autonomous Logistics Sorting — planning
  12. FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents — evaluation
  13. AgenticTwin: An Agentic LLM Framework Integrated with Digital Twin for Anomaly Detection — planning, engineering
  14. CLAIM: Leading Open-domain Active Clarification of Large Language Models with Uncertainty Measurement — other

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。