MAS Conference Papers：近期多智能体系统论文阅读清单
0. 一句话
这是一份近期 Multi-Agent Systems 论文的阅读优先级清单。重点不是收集最多条目，而是把和 Agent Swarm、协作结构、拓扑设计、runtime efficiency、过程验证最相关的论文排在前面。
1. 阅读顺序
我会优先读三类：
-
直接处理 MAS collaboration / topology / orchestration 的论文。
-
能作为系统 baseline 或 evaluation framework 的论文。
-
讨论效率、安全、bias、process verification 这类真实部署问题的论文。
2. 论文清单
| # | 论文 / 会议 | GitHub | arXiv | Rating | 我的判断 |
|---|---|---|---|---|---|
| 1 | Latent Collaboration in Multi-Agent Systems / ICML 2026 Spotlight (GitHub) | https://github.com/Gen-Verse/LatentMAS | https://arxiv.org/abs/2511.20639 | 9.4 | 最值得优先读。把多智能体协作放到 latent space 里做，直击 MAS 的通信/协作开销问题。 |
| 2 | MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems / ICML 2026 (GitHub) | https://github.com/wangzx1219/MASPO | https://arxiv.org/abs/2605.06623 | 9.1 | 面向 LLM-MAS 的联合 prompt 优化，问题定义很核心，代码公开，适合做 baseline。 |
| 3 | OMAC: A Holistic Optimization Framework for LLM-Based Multi-Agent Collaboration / ICML 2026 (arXiv) | 未找到稳定可信的官方 GitHub；作者页显示有 Codes | https://arxiv.org/abs/2505.11765 | 9.0 | 从 prompt、topology、communication 等整体优化 MAS，方向很正。录用标签在不同索引里有 Oral/Spotlight 差异，但“ICML 2026 收录”证据较强。 |
| 4 | Graph-of-Agents: Test-Time Scaling Laws via Collaboration of Heterogeneous Agents / ICLR 2026 (OpenReview) | https://github.com/UNITES-Lab/GoA | https://arxiv.org/abs/2604.17148 | 8.9 | “Agent graph + test-time scaling”非常贴近 Agent Swarm/异构协作，适合研究多 agent 组合结构。 |
| 5 | Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation / ARG-Designer / AAAI 2026 Oral (AAAI Publications) | https://github.com/Shiy-Li/ARG-Designer | https://arxiv.org/abs/2507.18224 | 8.8 | 自动设计 MAS 通信拓扑，和 swarm/crew 组织结构高度相关，AAAI Oral 加分。 |
| 6 | MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks / ICML 2026 (GitHub) | https://github.com/SalesforceAIResearch/MAS-Orchestra | https://arxiv.org/abs/2601.14652 | 8.8 | Salesforce 做的 MAS orchestration/benchmark，工程和评测价值高。 |
| 7 | Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs / ICLR 2026 (OpenReview) | https://github.com/pettingllms-ai/PettingLLMs | https://arxiv.org/abs/2510.11062 | 8.7 | 用 MARL 训练协作型 LLM agents，方法上比单纯 prompt-engineering 更进一步。 |
| 8 | MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks / ICML 2026 Spotlight (arXiv) | 未找到可信官方 GitHub | https://arxiv.org/abs/2603.02630 | 8.7 | Bandit + GNN 做 MAS prompt/topology 优化，理论味和系统味兼具。 |
| 9 | Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies / ICLR 2026 (OpenReview) | 未找到可信官方 GitHub | https://arxiv.org/abs/2502.02533 | 8.6 | 主题直接命中“prompt + topology”设计，是 MAS 结构优化方向的重要参考。 |
| 10 | MARTI: A Framework for Multi-Agent LLM Systems Reinforced Training and Inference / ICLR 2026 (OpenReview) | https://github.com/TsinghuaC3I/MARTI | https://arxiv.org/abs/2602.07848 | 8.5 | 清华系开源框架，覆盖多智能体强化训练和推理，实用性强。 |
| 11 | Stop Wasting Your Tokens: Towards Efficient Runtime Multi-Agent Systems / ICLR 2026 (GitHub) | https://github.com/LINs-lab/SupervisorAgent | https://arxiv.org/abs/2510.26585 | 8.3 | 解决 MAS runtime token 浪费，偏系统效率，对实际部署很有用。 |
| 12 | Benefits and Limitations of Communication in Multi-Agent Reasoning / ICLR 2026 (OpenReview) | https://github.com/michaelrizvi/coa-algorithmic | https://arxiv.org/abs/2510.13903 | 8.2 | 研究多 agent 通信到底什么时候有用/无用，适合做 MAS 机制分析。 |
| 13 | Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems / ICLR 2026 (GitHub) | https://github.com/weizhihao1/MAS-Bias | https://arxiv.org/abs/2604.08963 | 8.2 | 直接研究 “biased swarm”，是安全/对齐视角下的 MAS 重要论文。 |
| 14 | MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems / ICML 2026 (ICML) | https://github.com/Wang-ML-Lab/MAS-ProVe | https://arxiv.org/abs/2602.03053 | 8.1 | 面向 MAS 过程验证，适合关心 agent workflow 是否可靠的人。 |
| 15 | Which LLM Multi-Agent Protocol to Choose? / ICML 2026 (arXiv) | https://github.com/ulab-uiuc/AgentProtocols | https://arxiv.org/abs/2510.17149 | 7.9 | 对比不同 LLM multi-agent protocol，很适合快速了解协议选择和 benchmark。 |
| 16 | Agent Primitives: Reusable Latent Building Blocks for Multi-Agent Systems / ICML 2026 (ICML) | 未找到可信官方 GitHub | https://arxiv.org/abs/2602.03695 | 7.8 | “可复用 agent primitives”这个想法不错，但我没核到官方代码，扣分。 |
| 17 | AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation / ICML 2026 (ICML) | 未找到可信官方 GitHub | https://arxiv.org/abs/2602.17100 | 7.8 | 把 MAS topology evolution 用在竞赛级代码生成，应用明确，和拓扑搜索强相关。 |
| 18 | Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation / ACL Findings 2026 (GitHub) | https://github.com/NuoJohnChen/MAS_Diversity | https://arxiv.org/abs/2604.18005 | 7.8 | 研究多智能体系统里“集体失效/多样性塌缩”，对 swarm behavior 分析很有价值。 |
| 19 | A4VL: A Multi-Agent Perception-Action Alliance for Efficient Long Video Reasoning / CVPR 2026 (CVF Open Access) | https://github.com/git-disl/A4VL | https://arxiv.org/abs/2603.14052 | 7.7 | CV/视频方向的多智能体感知-行动协作框架，应用型强，通用 MAS 方法性稍弱。 |
| 20 | Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding / VideoHV-Agent / CVPR 2026 (CVF Open Access) | https://github.com/Haorane/VideoHV-Agent | https://arxiv.org/abs/2603.04977 | 7.6 | 多 agent 假设-验证式长视频理解，适合看 MAS 在 multimodal reasoning 里的落地。 |