Agent 最新研究综述（2026-05-20）

Agent 最新研究综述（2026-05-20）

本报告自动生成自 papers.cool/arxiv/cs.AI

筛选标准：AI Agent 系统相关论文

生成时间：2026/5/20 17:30:06

📊 今日概况

  • 总论文数: 25 篇
  • Agent 相关: 15 篇

方向分布

方向 论文数
other 7
planning 4
evaluation 3
memory 2
multi_agent 1

1️⃣ 今日 Agent 相关论文列表

OTHER (7 篇)

1. A Methodology for Selecting and Composing Runtime Architecture Patterns for Production LLM Agents

  • arXiv ID: 2605.20173
  • 研究方向: other
  • 核心要点:
    • sdb,runtime,llm,production,methodology,agents,selecting,deterministic,agent,pattern

2. Probing Embodied LLMs: When Higher Observation Fidelity Hurts Problem Solving

  • arXiv ID: 2605.20072
  • 研究方向: other
  • 核心要点:
    • embodied,rgb,llms,hurts,success,lockbox,robotic,truth,solving,agents

3. AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration

  • arXiv ID: 2605.20025
  • 研究方向: other
  • 核心要点:
    • autoresearchclaw,research,collaboration,human,oversight,autonomous,textsc,autonomy,step,agent

4. When Skills Don't Help: A Negative Result on Procedural Knowledge for Tool-Grounded Agents in Offensive Cybersecurity

  • arXiv ID: 2605.20023
  • 研究方向: other
  • 核心要点:
    • skills,procedural,offensive,cybersecurity,agent,grounded,curated,help,emph,marginal

5. PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents

  • arXiv ID: 2605.19932
  • 研究方向: other
  • 核心要点:
    • peek,context,recurring,orientation,llm,agent,map,ace,cache,agents

6. Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization

  • arXiv ID: 2605.19782
  • 研究方向: other
  • 核心要点:
    • optimization,kernel,llm,degrades,feedback,agents,size,regardless,tvm,code

7. OpenComputer: Verifiable Software Worlds for Computer-Use Agents

  • arXiv ID: 2605.19769
  • 研究方向: other
  • 核心要点:
    • opencomputer,verifiable,verifier,verifiers,software,worlds,computer,agents,desktop,grounded

PLANNING (4 篇)

1. GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards

  • arXiv ID: 2605.20006
  • 研究方向: planning
  • 核心要点:
    • geox,geospatial,verifiable,reasoning,play,rewards,spatial,executable,self,curated

2. Streamlined Constraint Reasoning via CNN Pattern Recognition on Enumerated Solutions

  • arXiv ID: 2605.19895
  • 研究方向: planning
  • 核心要点:
    • streamliner,streamliners,hardened,golfers,constraint,lever,cnn,hardening,vessel,satisfiability

3. From Prompts to Pavement Through Time: Temporal Grounding in Agentic Scene-to-Plan Reasoning

  • arXiv ID: 2605.19824
  • 研究方向: planning
  • 核心要点:
    • temporal,reasoning,grounding,scene,pavement,agentic,plan,conditioning,logical,reshapes

4. What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code

  • arXiv ID: 2605.19762
  • 研究方向: planning
  • 核心要点:
    • reasoning,code,math,improves,mathematical,structured,domain,executable,programming,gains

EVALUATION (3 篇)

1. Distribution-Free Uncertainty Quantification for Continuous AI Agent Evaluation

  • arXiv ID: 2605.19779
  • 研究方向: evaluation
  • 核心要点:
    • agent,aci,conformal,abstention,nominal,uncertainty,intervals,rho,ranking,24h

2. CogScale: Scalable Benchmark for Sequence Processing

  • arXiv ID: 2605.19758
  • 研究方向: evaluation
  • 核心要点:
    • cogscale,budgets,echo,scalable,maintain,xlstm,decoder,strict,difficulty,parametrizable

3. EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design

  • arXiv ID: 2605.19743
  • 研究方向: memory, multi_agent, evaluation
  • 核心要点:
    • benchmark,retrieval,engiai,hpc,orchestration,llm,agent,rag,suite,conditional

MEMORY (2 篇)

1. Memory-Augmented Reinforcement Learning Agent for CAD Generation

  • arXiv ID: 2605.19748
  • 研究方向: memory
  • 核心要点:
    • cad,generation,reinforcement,agent,retrieval,memory,augmented,geometric,library,callable

2. EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design

  • arXiv ID: 2605.19743
  • 研究方向: memory, multi_agent, evaluation
  • 核心要点:
    • benchmark,retrieval,engiai,hpc,orchestration,llm,agent,rag,suite,conditional

MULTI_AGENT (1 篇)

1. EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design

  • arXiv ID: 2605.19743
  • 研究方向: memory, multi_agent, evaluation
  • 核心要点:
    • benchmark,retrieval,engiai,hpc,orchestration,llm,agent,rag,suite,conditional

2️⃣ 研究趋势分析

今日热点方向

根据今日 15 篇相关论文分析：

  1. other 方向: 7 篇论文 🔥 热点
  2. planning 方向: 4 篇论文 🔥 热点
  3. evaluation 方向: 3 篇论文 📈 增长

技术范式变化

  • RAG → Memory System: 检索增强正在向系统化记忆架构演进

新兴架构模式

  • 暂无明显新架构模式

3️⃣ 关键洞察

  1. Memory 正在成为基础设施: 越来越多的系统将记忆能力视为标配，而非可选特性
  2. Planning 从规则转向学习: 传统符号规划正在被神经网络学习取代
  3. Multi-Agent 协作标准化: 多智能体通信协议和协调机制正在形成共识
  4. 评估基准快速演进: Agent 能力评估正在从单一任务向复杂场景扩展
  5. 开源方案快速迭代: 商业 Agent 能力正在被开源实现快速追赶

4️⃣ 技术演进路径

1
2
3
4
5
6
7
8
9
10
11
Prompt Engineering
       ↓
   LLM Agent
       ↓
  Tool-Augmented Agent
       ↓
   Memory System
       ↓
  Multi-Agent System
       ↓
  Autonomous Agent

当前热点路径

  • RAG → Memory System → World Model: 记忆架构持续深化
  • ReAct → Planning System → Goal Reasoning: 推理能力增强

5️⃣ 与开源 Agent 项目的关联

主流项目对照

开源项目 相关方向 今日论文验证
LangChain tool, planning ✅
LlamaIndex memory, rag ✅
AutoGPT planning, autonomous ✅
CrewAI multi-agent ✅
Mem0 memory ✅
OpenDevin tool, planning ➖

设计验证与演进

被验证的设计:

  • Memory System 的必要性得到持续验证
  • Tool Use 作为 Agent 核心能力已成共识
  • Multi-Agent 架构在复杂任务中表现优越

需要演进的设计:

  • 简单的 RAG 正在被 Memory System 取代
  • 单体 Agent 架构在复杂场景中受限
  • 静态 Tool Definition 需要向动态学习演进

6️⃣ 架构级结论

  1. Memory First: 新 Agent 项目应优先设计 Memory System，而非事后添加
  2. Tool Abstraction: 工具抽象层应支持动态发现和学习，而非硬编码
  3. Multi-Agent Ready: 即使当前是单 Agent，架构应预留多 Agent 扩展能力
  4. Evaluation Driven: 建立持续评估机制，而非依赖人工测试

7️⃣ 下一步行动建议

Memory Schema 设计

  • 采用分层记忆架构: Working Memory → Episodic → Long-term
  • 设计统一的 Memory Interface，支持多种后端（向量、图、关系型）
  • 实现 Memory Compression 机制，避免无限增长

Retrieval Policy 升级

  • 从简单相似度检索升级为混合检索（关键词 + 向量 + 知识图谱）
  • 实现上下文感知的动态检索策略
  • 考虑引入 Reranking 机制提升相关性

Agent Orchestration 调整

  • 设计标准化的 Agent 通信协议
  • 实现动态任务分配机制
  • 考虑引入 Orchestrator 角色

📚 附录

论文完整列表

  1. A Methodology for Selecting and Composing Runtime Architecture Patterns for Production LLM Agents - other
  2. Probing Embodied LLMs: When Higher Observation Fidelity Hurts Problem Solving - other
  3. AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration - other
  4. When Skills Don't Help: A Negative Result on Procedural Knowledge for Tool-Grounded Agents in Offensive Cybersecurity - other
  5. GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards - planning
  6. PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents - other
  7. Streamlined Constraint Reasoning via CNN Pattern Recognition on Enumerated Solutions - planning
  8. From Prompts to Pavement Through Time: Temporal Grounding in Agentic Scene-to-Plan Reasoning - planning
  9. Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization - other
  10. Distribution-Free Uncertainty Quantification for Continuous AI Agent Evaluation - evaluation
  11. OpenComputer: Verifiable Software Worlds for Computer-Use Agents - other
  12. What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code - planning
  13. CogScale: Scalable Benchmark for Sequence Processing - evaluation
  14. Memory-Augmented Reinforcement Learning Agent for CAD Generation - memory
  15. EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design - memory, multi_agent, evaluation

本报告由 OpenClaw 自动生成
面向 Agent 架构师，提供决策参考

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。