Agent 最新研究综述（2026-05-13）

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：2026/5/13 17:30:06

---

📊 今日概况

• 总论文数: 25 篇
• Agent 相关: 10 篇

方向分布

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

---

1️⃣ 今日 Agent 相关论文列表

OTHER (3 篇)

1. ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents

• arXiv ID: 2605.12481
• 研究方向: other
• 核心要点:
  • gui,tool,toolcua,path,orchestration,trajectories,agents,interleaved,trajectory,cuas

2. ProfiliTable: Profiling-Driven Tabular Data Processing via Agentic Workflows

• arXiv ID: 2605.12376
• 研究方向: other
• 核心要点:
  • profilitable,profiling,tabular,agentic,ambiguous,table,task,execution,refinement,workflows

3. Rollout Cards: A Reproducibility Standard for Agent Research

• arXiv ID: 2605.12131
• 研究方向: other
• 核心要点:
  • rollout,scores,cards,reporting,agent,reported,reproducibility,ergon,research,rules

---

PLANNING (4 篇)

1. Reinforcing VLAs in Task-Agnostic World Models

• arXiv ID: 2605.12334
• 研究方向: planning
• 核心要点:
  • vlas,world,task,reinforcing,agnostic,rollouts,reward,dream,vla,costly

2. Executable Agentic Memory for GUI Agent

• arXiv ID: 2605.12294
• 研究方向: memory, planning
• 核心要点:
  • gui,eam,agentic,executable,androidworld,memory,tars,horizon,planning,steers

3. Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM Systems

• arXiv ID: 2605.12213
• 研究方向: memory, planning
• 核心要点:
  • goal,reasoning,memory,mem,rag,agentic,retrieved,conversational,llm,hop

4. Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics

• arXiv ID: 2605.12178
• 研究方向: planning
• 核心要点:
  • enterprise,dynamics,world,internalized,agents,runtime,configurable,shift,discovery,grounding

---

MEMORY (3 篇)

1. Executable Agentic Memory for GUI Agent

• arXiv ID: 2605.12294
• 研究方向: memory, planning
• 核心要点:
  • gui,eam,agentic,executable,androidworld,memory,tars,horizon,planning,steers

2. Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM Systems

• arXiv ID: 2605.12213
• 研究方向: memory, planning
• 核心要点:
  • goal,reasoning,memory,mem,rag,agentic,retrieved,conversational,llm,hop

3. MolDeTox: Evaluating Language Model's Stepwise Fragment Editing for Molecular Detoxification

• arXiv ID: 2605.12181
• 研究方向: memory
• 核心要点:
  • moldetox,toxicity,detoxification,stepwise,molecular,fragment,vlms,molecules,language,drug

---

MULTI_AGENT (1 篇)

1. No Action Without a NOD: A Heterogeneous Multi-Agent Architecture for Reliable Service Agents

• arXiv ID: 2605.12240
• 研究方向: multi_agent
• 核心要点:
  • nod,service,agents,navigator,director,agent,hallucinations,violations,heterogeneous,task

---

EVALUATION (1 篇)

1. MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling

• arXiv ID: 2605.12154
• 研究方向: evaluation
• 核心要点:
  • solver,optbench,optimization,multimodal,mllms,instances,grounded,modeling,780,executable

---

2️⃣ 研究趋势分析

今日热点方向

根据今日 10 篇相关论文分析：

1. planning 方向: 4 篇论文 🔥 热点
2. other 方向: 3 篇论文 📈 增长
3. memory 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
• Tool Calling → Tool Learning: 从简单工具调用到自主工具学习

新兴架构模式

• Agent Workflow: 工作流编排架构

---

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

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. ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents - other
2. ProfiliTable: Profiling-Driven Tabular Data Processing via Agentic Workflows - other
3. Reinforcing VLAs in Task-Agnostic World Models - planning
4. Executable Agentic Memory for GUI Agent - memory, planning
5. No Action Without a NOD: A Heterogeneous Multi-Agent Architecture for Reliable Service Agents - multi_agent
6. Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM Systems - memory, planning
7. MolDeTox: Evaluating Language Model's Stepwise Fragment Editing for Molecular Detoxification - memory
8. Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics - planning
9. MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling - evaluation
10. Rollout Cards: A Reproducibility Standard for Agent Research - other

---

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