数据来源：papers.cool/arxiv/cs.AI · 生成时间：2026/11/6 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 22 篇论文，覆盖 2 天数据。上周 45 篇，环比减少 23 篇。
研究方向分布
| 方向 | 本周 | 上周 | 变化 |
|---|---|---|---|
| 其他 | 7 | 19 | -12 |
| 记忆系统 | 6 | 10 | -4 |
| 评估基准 | 4 | 6 | -2 |
| 规划推理 | 3 | 7 | -4 |
| 安全对齐 | 3 | 2 | +1 |
| 工程架构 | 2 | 2 | 0 |
| 自我进化 | 1 | 1 | 0 |
应用场景分布
| 场景 | 论文数 | 占比 |
|---|---|---|
| 创意与内容 | 2 | 9% |
| 机器人与物理世界 | 2 | 9% |
| 企业自动化 | 2 | 9% |
| 信息检索与问答 | 2 | 9% |
| 决策支持 | 1 | 5% |
| 科学研究 | 1 | 5% |
| 代码开发 | 1 | 5% |
核心论文解读
1. FrontierPhysics: Evaluating Agents for End-to-End Physics Research
- 英文标题: FrontierPhysics: Evaluating Agents for End-to-End Physics Research
- arXiv: 2611.06910 Kimi解读
- 方向: 其他
- 场景: 科学研究、企业自动化、信息检索与问答
2. Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling
- 英文标题: Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling
- arXiv: 2610.41909 Kimi解读
- 方向: 记忆系统 · 工程架构
- 场景: 决策支持
3. BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories
- 英文标题: BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories
- arXiv: 2611.06911 Kimi解读
- 方向: 安全对齐 · 评估基准
- 场景: 机器人与物理世界
4. TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents
- 英文标题: TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents
- arXiv: 2610.42004 Kimi解读
- 方向: 评估基准 · 自我进化
5. Learning to Interact and Direct: Co-Evolving Video Director Agents
- 英文标题: Learning to Interact and Direct: Co-Evolving Video Director Agents
- arXiv: 2610.41865 Kimi解读
- 方向: 其他
- 场景: 创意与内容
6. CompressionBench: How can your coding agents compress LLMs?
- 英文标题: CompressionBench: How can your coding agents compress LLMs?
- arXiv: 2611.06858 Kimi解读
- 方向: 记忆系统
- 场景: 企业自动化
7. ServerBench: How can your coding agents design specialized LLM serving systems?
- 英文标题: ServerBench: How can your coding agents design specialized LLM serving systems?
- arXiv: 2611.06841 Kimi解读
- 方向: 其他
- 场景: 创意与内容
8. ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
- 英文标题: ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
- arXiv: 2611.06732 Kimi解读
- 方向: 规划推理
- 场景: 代码开发
9. Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification
- 英文标题: Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification
- arXiv: 2611.06662 Kimi解读
- 方向: 规划推理
- 场景: 信息检索与问答
10. Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models
- 英文标题: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models
- arXiv: 2611.06553 Kimi解读
- 方向: 记忆系统 · 安全对齐
研究趋势
主导方向：其他（7 篇），较上周（19 篇）下降。
上升: 安全对齐（2→3）
下降: 记忆系统（10→6）、其他（19→7）、多智能体（3→0）、规划推理（7→3）、评估基准（6→4）
技术演进脉络
其他（7 篇）
- Optimizing Stochastic Policy Geometry for Few-Step Flow Reinforcement Learning Kimi解读
- Learning to Interact and Direct: Co-Evolving Video Director Agents Kimi解读
- FrontierPhysics: Evaluating Agents for End-to-End Physics Research Kimi解读
- 及另外 4 篇
记忆系统（6 篇）
- DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention Kimi解读
- Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling Kimi解读
- CompressionBench: How can your coding agents compress LLMs? Kimi解读
- 及另外 3 篇
评估基准（4 篇）
- TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents Kimi解读
- BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories Kimi解读
- Task-Efficient Agent Evaluation Using Cached Traces Kimi解读
- 及另外 1 篇
规划推理（3 篇）
- RECAST: REconstruction of Chain-of-Thought via Auditable Structured Traces Kimi解读
- ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models Kimi解读
- Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification Kimi解读
安全对齐（3 篇）
- Whose Values, Which Contexts, and What Signals? Grounding Cultural Value Alignment with Re3Culture Kimi解读
- BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories Kimi解读
- Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models Kimi解读
工程架构（2 篇）
- Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling Kimi解读
- EviSeek: An Evidence-Centric Agentic Framework for Embodied Active Information Seeking Kimi解读
自我进化（1 篇）
工程实践启示
- 工程架构方向 2 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。
- 安全方向 3 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 7 篇，上周 19 篇）、记忆系统（本周 6 篇，上周 10 篇）、评估基准（本周 4 篇，上周 6 篇）、规划推理（本周 3 篇，上周 7 篇）、安全对齐（本周 3 篇，上周 2 篇）、工程架构（本周 2 篇，上周 2 篇）
附录：本周论文完整列表
去重后共 22 篇。
2026-11-02（7 篇）
- TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents Kimi解读 — evaluation, evolution
- DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention Kimi解读 — memory
- Optimizing Stochastic Policy Geometry for Few-Step Flow Reinforcement Learning Kimi解读 — other
- Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling Kimi解读 — memory, engineering
- Learning to Interact and Direct: Co-Evolving Video Director Agents Kimi解读 — other
- RECAST: REconstruction of Chain-of-Thought via Auditable Structured Traces Kimi解读 — planning
- Whose Values, Which Contexts, and What Signals? Grounding Cultural Value Alignment with Re3Culture Kimi解读 — safety
2026-11-05（15 篇）
- BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories Kimi解读 — safety, evaluation
- FrontierPhysics: Evaluating Agents for End-to-End Physics Research Kimi解读 — other
- On Turn-Level Best-of-N Sampling for Scaling LLM Agents Kimi解读 — other
- CompressionBench: How can your coding agents compress LLMs? Kimi解读 — memory
- ServerBench: How can your coding agents design specialized LLM serving systems? Kimi解读 — other
- Task-Efficient Agent Evaluation Using Cached Traces Kimi解读 — evaluation
- ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models Kimi解读 — planning
- LifeVal: Evaluating AI Model Performance on Personalized Agentic Consumer Tasks Kimi解读 — other
- Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification Kimi解读 — planning
- RAGAM-AI: A Hybrid Reinforcement Learning based Agentic approach powered by Generative AI for Edge Resource Management Kimi解读 — memory
- Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models Kimi解读 — memory, safety
- Tools Need Not Speak in Text: Latent Context Compression for Agents Kimi解读 — other
- AmbiHTMLBench: A Benchmark for HTML-based Visually Rich Document parsing Kimi解读 — evaluation
- SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory Kimi解读 — memory
- EviSeek: An Evidence-Centric Agentic Framework for Embodied Active Information Seeking Kimi解读 — engineering
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。