Agent 研究周度深度综述（2026-09-14 ~ 2026-09-20）

数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/9/18 17:00:04
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 53 篇，环比增加 2 篇。

研究方向分布

方向 本周 上周 变化
规划推理 18 17 +1
其他 15 10 +5
记忆系统 10 10 0
评估基准 7 13 -6
工程架构 7 10 -3
安全对齐 4 6 -2
自我进化 4 1 +3
多智能体 3 4 -1

应用场景分布

场景 论文数 占比
信息检索与问答 12 22%
科学研究 7 13%
决策支持 4 7%
企业自动化 3 5%
机器人与物理世界 3 5%
代码开发 2 4%
数据分析 1 2%

核心论文解读

1. Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrieval

  • 英文标题: Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrieval
  • arXiv: 2609.13073 Kimi解读
  • 方向: 记忆系统 · 工程架构
  • 场景: 科学研究、信息检索与问答

2. Embodied-BenchForge: A Closed-Loop Agentic Workflow for Embodied Benchmark Construction

  • 英文标题: Embodied-BenchForge: A Closed-Loop Agentic Workflow for Embodied Benchmark Construction
  • arXiv: 2609.13082 Kimi解读
  • 方向: 评估基准
  • 场景: 企业自动化、机器人与物理世界

3. Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science

  • 英文标题: Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science
  • arXiv: 2609.15983 Kimi解读
  • 方向: 其他
  • 场景: 科学研究、信息检索与问答

4. AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery

  • 英文标题: AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery
  • arXiv: 2609.15820 Kimi解读
  • 方向: 其他
  • 场景: 科学研究、信息检索与问答

5. Atria Dawn: The Dawn of Agentic Superintelligence

  • 英文标题: Atria Dawn: The Dawn of Agentic Superintelligence
  • arXiv: 2609.15818 Kimi解读
  • 方向: 其他
  • 场景: 科学研究、信息检索与问答

6. ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents

  • 英文标题: ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
  • arXiv: 2609.17523 Kimi解读
  • 方向: 自我进化
  • 场景: 科学研究、信息检索与问答

7. Transformer-Based Token Fusion and Dynamic Graph Planning for Audio-Visual Navigation

  • 英文标题: Transformer-Based Token Fusion and Dynamic Graph Planning for Audio-Visual Navigation
  • arXiv: 2609.17421 Kimi解读
  • 方向: 规划推理
  • 场景: 决策支持、机器人与物理世界

8. FlashVector: Agent for Hierarchical Model Serving Stack Optimization

  • 英文标题: FlashVector: Agent for Hierarchical Model Serving Stack Optimization
  • arXiv: 2609.17391 Kimi解读
  • 方向: 工程架构
  • 场景: 代码开发、决策支持

9. Intrinsic Motivation in Reinforcement Learning: A Research Agenda for Adaptive Self-Organisation

  • 英文标题: Intrinsic Motivation in Reinforcement Learning: A Research Agenda for Adaptive Self-Organisation
  • arXiv: 2609.17325 Kimi解读
  • 方向: 自我进化
  • 场景: 科学研究、信息检索与问答

10. Symbolic Separation: Grounding Deep Agents in Knowledge Graphs for Trustworthy Operational Data Analytics

  • 英文标题: Symbolic Separation: Grounding Deep Agents in Knowledge Graphs for Trustworthy Operational Data Analytics
  • arXiv: 2609.17107 Kimi解读
  • 方向: 其他
  • 场景: 数据分析、信息检索与问答

研究趋势

主导方向：规划推理（18 篇），较上周（17 篇）上升。

上升: 规划推理（17→18）、其他（10→15）、自我进化（1→4）

下降: 工具使用（2→0）、评估基准（13→7）、工程架构（10→7）、多智能体（4→3）、安全对齐（6→4）

技术演进脉络

规划推理（18 篇）

其他（15 篇）

记忆系统（10 篇）

评估基准（7 篇）

工程架构（7 篇）

安全对齐（4 篇）

自我进化（4 篇）

多智能体（3 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。
  • 记忆系统方向 10 篇，RAG 与长期记忆方案不断优化。
  • 多智能体方向 3 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：规划推理（本周 18 篇，上周 17 篇）、其他（本周 15 篇，上周 10 篇）、记忆系统（本周 10 篇，上周 10 篇）、评估基准（本周 7 篇，上周 13 篇）、工程架构（本周 7 篇，上周 10 篇）、安全对齐（本周 4 篇，上周 6 篇）、多智能体（本周 3 篇，上周 4 篇）

附录：本周论文完整列表

去重后共 55 篇。

2026-09-14（8 篇）

  1. CMA-OT: Hierarchical Expert Supervision for Dance-to-Music Generation Kimi解读 — safety
  2. Embodied-BenchForge: A Closed-Loop Agentic Workflow for Embodied Benchmark Construction Kimi解读 — evaluation
  3. Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrieval Kimi解读 — memory, engineering
  4. How Good Are Frontier Models at Physics? Expert Re-Grading Reveals Broken Evaluations and Near-Saturation of Leading Benchmarks Kimi解读 — evaluation
  5. K-Bench: A Benchmark for LLM Unlearning in Agentic Deployments Kimi解读 — evaluation, engineering
  6. Unified Agentic Video Editing Across Levels of Complexity and Creativity Kimi解读 — other
  7. What Drives Recovery in Agentic Text-to-Cypher? LAST-CQ: An LLM Agent Self-Refinement Framework Kimi解读 — engineering
  8. Skill Issue: Lessons from Optimizing Repository SKILLs for Coding Agents Kimi解读 — other

2026-09-15（15 篇）

  1. Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection Kimi解读 — planning
  2. Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science Kimi解读 — other
  3. Recurrent GraphNeural NetworkswithSet-BasedAggregation Kimi解读 — memory
  4. LongAgent: History-Guided Agentic Search for Longitudinal Outcome Prediction Kimi解读 — other
  5. AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery Kimi解读 — other
  6. Atria Dawn: The Dawn of Agentic Superintelligence Kimi解读 — other
  7. When Should a World Model Move? Loss-Conditioned State Execution Kimi解读 — planning
  8. Navigating Sparse Evidence: Agentic Visual RAG via Explicit Context Selection and Consolidation Kimi解读 — memory, planning
  9. KnowBench: Effort Reduction as a Unified, Deployment-Grounded Benchmark for Clinical AI Kimi解读 — evaluation, engineering
  10. EvoOntology: A Self-Evolving Ontology Layer for Data Agents Kimi解读 — other
  11. NoteVQA: Benchmarking VLMs on Real-Life Questions from Human Communities Kimi解读 — evaluation
  12. Diversified and Perceptible Counterfactual Examples Leveraging Expert Knowledge Kimi解读 — memory
  13. Option-Aware Retrieval and Task-Specific VLM Adaptation for Medical VQA Kimi解读 — memory
  14. Beyond Safe Answers: Segment-Aware Listwise Alignment for Reasoning Safety in Large Reasoning Models Kimi解读 — planning, safety
  15. The Troy Moment of AI: Why SomeWill Cheat and SomeWill Follow? Kimi解读 — other

2026-09-16（15 篇）

  1. ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents Kimi解读 — evolution
  2. Verifiable Social Reasoning for LLM Assistants Kimi解读 — planning
  3. Transformer-Based Token Fusion and Dynamic Graph Planning for Audio-Visual Navigation Kimi解读 — planning
  4. World Model Science: Self-Organized Criticality, Weak Chaos, and Metastable Belief Dynamics in Long-Horizon LLM Agents Kimi解读 — planning
  5. Never Stop Thinking: Continuous-Time Language Agents Kimi解读 — other
  6. FlashVector: Agent for Hierarchical Model Serving Stack Optimization Kimi解读 — engineering
  7. Self-Emergence Agent Architecture:Behavior-Inertia HMM, Reflexive Metacognition,and Social-Contrastive Self-Modeling Kimi解读 — engineering
  8. Intrinsic Motivation in Reinforcement Learning: A Research Agenda for Adaptive Self-Organisation Kimi解读 — evolution
  9. End-to-End Latency-Minimizing and Load-Balanced Request Scheduling for Edge LLM Inference in Agentic AI Services Kimi解读 — other
  10. MOCC-R1: Reinforcing Reasoning-Response Consistency for Multimodal Counselor Response Generation Kimi解读 — planning
  11. FirmCORe: A Benchmark for Structured Reasoning about Inter-Firm Collaboration Opportunities Kimi解读 — planning, evaluation
  12. Symbolic Separation: Grounding Deep Agents in Knowledge Graphs for Trustworthy Operational Data Analytics Kimi解读 — other
  13. Scaling-Score Conformal Prediction for Multi-Target Regression Kimi解读 — memory
  14. Interactive Memory Learning for Long-Term Conversations Kimi解读 — memory
  15. SKIP: a Self-knowledge-guided Step-wise Preference Learning Framework for Concise Reasoning Kimi解读 — planning, engineering

2026-09-17（17 篇）

  1. Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments Kimi解读 — memory, evolution
  2. Flag Game: A Toy Model for Mechanistic Swarm Interpretability Kimi解读 — multi_agent
  3. MUSE: Benchmarking Large Vision-Language Models on Multi-Modal Understanding in Situated Education Kimi解读 — evaluation
  4. Compiled Agency: Frontier General-Purpose Coding Agents Build Winning Game Players from Bare Interaction - from Flappy Bird to StarCraft II and Civilization Kimi解读 — other
  5. Lost in Perception: Isolating Perceptual and Reasoning Failures in Multimodal Physics and Geometry Reasoning Kimi解读 — planning
  6. Compositional Policy Violations: When Step-Level Compliance Fails In Agentic AI Workflows Kimi解读 — other
  7. CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents Kimi解读 — other
  8. Clueing up LLMs with Tool-Augmented Deductive Reasoning Kimi解读 — planning
  9. Which LLM is Best for Translating Natural Language Goals to PDDL Kimi解读 — planning
  10. Reasoning through Evolution: Automatic Meta-path Discovery for LLM-based Fake News Detection Kimi解读 — planning
  11. Recursive Reasoning or Statistical Extrapolation? In-Context Learning in Multi-Agent Interdependent Decision-Making Kimi解读 — planning, multi_agent
  12. AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Execution Kimi解读 — multi_agent, evolution
  13. Beyond Routine Compliance: Cunning Data Cultivates Safety Vigilance in Large Language Models Kimi解读 — planning, safety
  14. First Token Matters: Understanding Safety Collapse in Large Reasoning Models Kimi解读 — planning, safety
  15. Disentangling Long-Term Memory via Latent Neuro-Symbolic Reasoning Kimi解读 — memory, planning
  16. Collective Loss of Control in LLM Agent Systems: An Epidemic Account of Mutation, Contagion, and Recovery Kimi解读 — other
  17. The Mirage of Calibrated Confidence: Trajectory-Independence of Verbalized Confidence in Vision-Language Models Kimi解读 — memory

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。