{ "schema_version": "1.0.0", "evidence_pack_id": "age-174-agent-teams-2026", "manifest_version": "2.0.0", "snapshot_digest": "sha256:1cab26e51999310225fb08e05621ddfdbcad7ca3e478bc37181af0d614484a8c", "digest_method": "SHA-256 of recursively key-sorted evidence payload after omitting top-level snapshot_digest, stable_url, evidence_snapshot_url, evidence_snapshot, validation.manifest_consumer_trial, and validation.consumer_attestation", "supersedes": [ { "manifest_version": "1.0.0", "snapshot_digest": "sha256:194c019808da705ac100cccd215155c5b09f67a86f9499abb0d808ab2a855170", "immutable_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/a7af1b76ac31a216935b74af14593da9fef0cee5/site/reports/agent-teams-2026/data/evidence-manifest.v1.json" } ], "request": { "request_id": "AGE-174", "questions": [ "LLM Agent Team 的角色分工、通信、协调、共享记忆、计划分解、批判/辩论、动态组队与规模效应有哪些相对可靠的结论？", "近 18 个月有哪些值得优先跟进的新机制、评测与失败研究？", "这些证据对 Multica Agent Team 的产品与架构设计有什么可执行启示？" ], "audience": "Multica workspace owner、Agent Team 产品与工程负责人", "scope": "以可通信、可行动、有持续状态的 LLM agent team 为主；覆盖 2023–2026 奠基论文与 2025-01-15 至 2026-07-15 前沿研究", "knowledge_cutoff": "2026-07-15", "retrieved_on": "2026-07-15", "time_precision": "day" }, "retrieval": [ { "query_id": "Q01", "query": "LLM multi-agent collaboration", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "角色、组织与通用协作", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q02", "query": "large language model multi-agent communication", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "消息协议、拓扑与 latent communication", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q03", "query": "large language model agent debate consensus", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "辩论、异构模型、从众与错误共识", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q04", "query": "LLM agent dynamic team routing", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "动态选人、架构搜索、路由与早停", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q05", "query": "multi-agent LLM benchmark collaboration", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "协作过程评测与失败 taxonomy", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q06", "query": "LLM multi-agent shared memory", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "共享状态、文件系统与权限化记忆", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q07", "query": "LLM agent orchestration planning", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "SOP、计划、工件与编排", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q08", "query": "LLM multi-agent safety failure", "source": "official arXiv and venue pages (oo unavailable; arxiv-cli plus primary-source fallback)", "coverage_note": "错误传播、终止、验证与安全", "run_on": "2026-07-15", "time_precision": "day" }, { "query_id": "Q09", "query": "multi-agent expert dilution reveal expert consensus", "source": "official arXiv and OpenReview", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "专家稀释与共识压力" }, { "query_id": "Q10", "query": "single agent multi-agent equal thinking token budget", "source": "official arXiv", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "matched-budget 强单体基线" }, { "query_id": "Q11", "query": "single agent baseline multi-agent workflow homogeneous collapse", "source": "official arXiv", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "同质 workflow 可折叠边界" }, { "query_id": "Q12", "query": "zero-cost collaboration capability protocol incentive LLM agents", "source": "official arXiv and ICML", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "能力与合作脱钩、协议和微激励" }, { "query_id": "Q13", "query": "distributed coordination communication reasoning gap multi-agent LLM", "source": "official arXiv and ACL Anthology", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "分布式信息整合与 Communication–Reasoning Gap" }, { "query_id": "Q14", "query": "HiddenBench hidden profile collective reasoning multi-agent", "source": "official arXiv", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "非对称信息下的集体推理失败" }, { "query_id": "Q15", "query": "LatentMAS learned alignment matrix identity no transfer", "source": "OpenReview CompLearn 2026", "run_on": "2026-07-15", "time_precision": "day", "coverage_note": "LatentMAS 直接反证" } ], "sources": [ { "source_id": "S01", "paper_id": "2303.17760", "title": "CAMEL: Communicative Agents for 'Mind' Exploration of Large Scale Language Model Society", "authors": [ "Guohao Li", "Hasan Abed Al Kader Hammoud", "Hani Itani", "Dmitrii Khizbullin", "Bernard Ghanem" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2303.17760v2", "venue_url": "https://proceedings.neurips.cc/paper/2023/hash/a3621ee907def47c1b952ade25c67698-Abstract-Conference.html", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2023, "venue": "NeurIPS", "track": "Conference" }, { "source_id": "S02", "paper_id": "2305.14325", "title": "Improving Factuality and Reasoning in Language Models through Multiagent Debate", "authors": [ "Yilun Du", "Shuang Li", "Antonio Torralba", "Joshua B. Tenenbaum", "Igor Mordatch" ], "version": "v1", "publication_status": "published", "official_url": "https://arxiv.org/abs/2305.14325v1", "venue_url": "https://proceedings.mlr.press/v235/du24e.html", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "ICML", "track": "Conference" }, { "source_id": "S03", "paper_id": "2307.07924", "title": "ChatDev: Communicative Agents for Software Development", "authors": [ "Chen Qian", "Wei Liu", "Hongzhang Liu", "Nuo Chen", "Yufan Dang", "Jiahao Li", "Cheng Yang", "Weize Chen", "Yusheng Su", "Xin Cong", "Juyuan Xu", "Dahai Li", "Zhiyuan Liu", "Maosong Sun" ], "version": "v5", "publication_status": "published", "official_url": "https://arxiv.org/abs/2307.07924v5", "venue_url": "https://aclanthology.org/2024.acl-long.810/", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "ACL", "track": "Long Paper" }, { "source_id": "S04", "paper_id": "2308.00352", "title": "MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework", "authors": [ "Sirui Hong", "Mingchen Zhuge", "Jiaqi Chen", "Xiawu Zheng", "Yuheng Cheng", "Ceyao Zhang", "Jinlin Wang", "Zili Wang", "Steven Ka Shing Yau", "Zijuan Lin", "Liyang Zhou", "Chenyu Ran", "Lingfeng Xiao", "Chenglin Wu", "Jürgen Schmidhuber" ], "version": "v7", "publication_status": "published", "official_url": "https://arxiv.org/abs/2308.00352v7", "venue_url": "https://openreview.net/forum?id=VtmBAGCN7o", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "ICLR", "track": "Oral" }, { "source_id": "S05", "paper_id": "2308.08155", "title": "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation", "authors": [ "Qingyun Wu", "Gagan Bansal", "Jieyu Zhang", "Yiran Wu", "Beibin Li", "Erkang Zhu", "Li Jiang", "Xiaoyun Zhang", "Shaokun Zhang", "Jiale Liu", "Ahmed Hassan Awadallah", "Ryen W White", "Doug Burger", "Chi Wang" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2308.08155v2", "venue_url": "https://openreview.net/forum?id=BAakY1hNKS", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "COLM", "track": "Conference" }, { "source_id": "S06", "paper_id": "2309.13007", "title": "ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs", "authors": [ "Justin Chih-Yao Chen", "Swarnadeep Saha", "Mohit Bansal" ], "version": "v3", "publication_status": "published", "official_url": "https://arxiv.org/abs/2309.13007v3", "venue_url": "https://aclanthology.org/2024.acl-long.381/", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "ACL", "track": "Long Paper" }, { "source_id": "S07", "paper_id": "2310.02170", "title": "A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration", "authors": [ "Zijun Liu", "Yanzhe Zhang", "Peng Li", "Yang Liu", "Diyi Yang" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2310.02170v2", "venue_url": "https://openreview.net/forum?id=XII0Wp1XA9", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "COLM", "track": "Conference" }, { "source_id": "S08", "paper_id": "2402.05120", "title": "More Agents Is All You Need", "authors": [ "Junyou Li", "Qin Zhang", "Yangbin Yu", "Qiang Fu", "Deheng Ye" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2402.05120v2", "venue_url": "https://openreview.net/forum?id=bgzUSZ8aeg", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2024, "venue_year": 2024, "venue": "TMLR", "track": "Journal" }, { "source_id": "S09", "paper_id": "2406.07155", "title": "Scaling Large Language Model-based Multi-Agent Collaboration", "authors": [ "Chen Qian", "Zihao Xie", "YiFei Wang", "Wei Liu", "Kunlun Zhu", "Hanchen Xia", "Yufan Dang", "Zhuoyun Du", "Weize Chen", "Cheng Yang", "Zhiyuan Liu", "Maosong Sun" ], "version": "v3", "publication_status": "published", "official_url": "https://arxiv.org/abs/2406.07155v3", "venue_url": "https://openreview.net/forum?id=K3n5jPkrU6", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2024, "venue_year": 2025, "venue": "ICLR", "track": "Conference" }, { "source_id": "S10", "paper_id": "2502.04180", "title": "Multi-agent Architecture Search via Agentic Supernet", "authors": [ "Guibin Zhang", "Luyang Niu", "Junfeng Fang", "Kun Wang", "Lei Bai", "Xiang Wang" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2502.04180v2", "venue_url": "https://openreview.net/forum?id=imcyVlzpXh", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": 2025, "venue": "ICML", "track": "Conference" }, { "source_id": "S11", "paper_id": "2502.11133", "title": "MasRouter: Learning to Route LLMs for Multi-Agent Systems", "authors": [ "Yanwei Yue", "Guibin Zhang", "Boyang Liu", "Guancheng Wan", "Kun Wang", "Dawei Cheng", "Yiyan Qi" ], "version": "v1", "publication_status": "preprint", "official_url": "https://arxiv.org/abs/2502.11133v1", "venue_url": null, "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": null, "venue": null, "track": null }, { "source_id": "S12", "paper_id": "2503.01935", "title": "MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents", "authors": [ "Kunlun Zhu", "Hongyi Du", "Zhaochen Hong", "Xiaocheng Yang", "Shuyi Guo", "Zhe Wang", "Zhenhailong Wang", "Cheng Qian", "Xiangru Tang", "Heng Ji", "Jiaxuan You" ], "version": "v1", "publication_status": "published", "official_url": "https://arxiv.org/abs/2503.01935v1", "venue_url": "https://aclanthology.org/2025.acl-long.421/", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": 2025, "venue": "ACL", "track": "Long Paper" }, { "source_id": "S13", "paper_id": "2503.13657", "title": "Why Do Multi-Agent LLM Systems Fail?", "authors": [ "Mert Cemri", "Melissa Z. Pan", "Shuyi Yang", "Lakshya A. Agrawal", "Bhavya Chopra", "Rishabh Tiwari", "Kurt Keutzer", "Aditya Parameswaran", "Dan Klein", "Kannan Ramchandran", "Matei Zaharia", "Joseph E. Gonzalez", "Ion Stoica" ], "version": "v3", "publication_status": "published", "official_url": "https://arxiv.org/abs/2503.13657v3", "venue_url": "https://proceedings.neurips.cc/paper_files/paper/2025/hash/b1041e52d3be19f0a9bc491657488e4a-Abstract-Datasets_and_Benchmarks_Track.html", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": 2025, "venue": "NeurIPS", "track": "Datasets and Benchmarks Track" }, { "source_id": "S14", "paper_id": "2505.21471", "title": "Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration", "authors": [ "Zijun Liu", "Zhennan Wan", "Peng Li", "Ming Yan", "Fei Huang", "Yang Liu" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2505.21471v2", "venue_url": "https://aclanthology.org/2026.acl-long.468/", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": 2026, "venue": "ACL", "track": "Long Paper" }, { "source_id": "S15", "paper_id": "2511.20639", "title": "Latent Collaboration in Multi-Agent Systems", "authors": [ "Jiaru Zou", "Ruizhong Qiu", "Gaotang Li", "Xiyuan Yang", "Katherine Tieu", "Pan Lu", "Ke Shen", "Hanghang Tong", "Yejin Choi", "Jingrui He", "James Zou", "Mengdi Wang", "Ling Yang" ], "version": "v3", "publication_status": "accepted", "official_url": "https://arxiv.org/abs/2511.20639v3", "venue_url": "https://openreview.net/forum?id=e7pAjJZJWb", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": 2026, "venue": "ICML", "track": "Spotlight" }, { "source_id": "S16", "paper_id": "2512.08296", "title": "Towards a Science of Scaling Agent Systems", "authors": [ "Yubin Kim", "Ken Gu", "Chanwoo Park", "Chunjong Park", "Samuel Schmidgall", "A. Ali Heydari", "Yao Yan", "Zhihan Zhang", "Yuchen Zhuang", "Yun Liu", "Mark Malhotra", "Paul Pu Liang", "Hae Won Park", "Yuzhe Yang", "Xuhai Xu", "Yilun Du", "Shwetak Patel", "Tim Althoff", "Daniel McDuff", "Xin Liu" ], "version": "v3", "publication_status": "preprint", "official_url": "https://arxiv.org/abs/2512.08296v3", "venue_url": null, "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2025, "venue_year": null, "venue": null, "track": null }, { "source_id": "S17", "paper_id": "2602.03794", "title": "Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity", "authors": [ "Yingxuan Yang", "Chengrui Qu", "Muning Wen", "Laixi Shi", "Ying Wen", "Weinan Zhang", "Adam Wierman", "Shangding Gu" ], "version": "v1", "publication_status": "workshop", "official_url": "https://arxiv.org/abs/2602.03794v1", "venue_url": "https://openreview.net/forum?id=9BN2W5BCfE", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2026, "venue_year": 2026, "venue": "ICLR", "track": "AIMS Workshop" }, { "source_id": "S18", "paper_id": "2602.01566", "title": "FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents", "authors": [ "Chiwei Zhu", "Benfeng Xu", "Mingxuan Du", "Shaohan Wang", "Xiaorui Wang", "Zhendong Mao", "Yongdong Zhang" ], "version": "v2", "publication_status": "published", "official_url": "https://arxiv.org/abs/2602.01566v2", "venue_url": "https://aclanthology.org/2026.acl-long.288/", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2026, "venue_year": 2026, "venue": "ACL", "track": "Long Paper" }, { "source_id": "S19", "paper_id": "2308.10848", "title": "AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors", "authors": [ "Weize Chen", "Yusheng Su", "Jingwei Zuo", "Cheng Yang", "Chenfei Yuan", "Chi-Min Chan", "Heyang Yu", "Yaxi Lu", "Yi-Hsin Hung", "Chen Qian", "Yujia Qin", "Xin Cong", "Ruobing Xie", "Zhiyuan Liu", "Maosong Sun", "Jie Zhou" ], "version": "v3", "publication_status": "published", "official_url": "https://arxiv.org/abs/2308.10848v3", "venue_url": "https://openreview.net/forum?id=EHg5GDnyq1", "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2023, "venue_year": 2024, "venue": "ICLR", "track": "Conference" }, { "source_id": "S20", "paper_id": "2606.18829", "title": "GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents", "authors": [ "Zhe Ren", "Yibo Yang", "Yimeng Chen", "Zijun Zhao", "Benshuo Fu", "Zhihao Shu", "Bingjie Zhang", "Yangyang Xu", "Dandan Guo", "Shuicheng Yan" ], "version": "v1", "publication_status": "preprint", "official_url": "https://arxiv.org/abs/2606.18829v1", "venue_url": null, "accessed_at": "2026-07-15", "source_kind": "arxiv", "submission_year": 2026, "venue_year": null, "venue": null, "track": null }, { "source_id": "S21", "source_kind": "arxiv", "paper_id": "2602.01011", "title": "Multi-Agent Teams Hold Experts Back", "authors": [ "Aneesh Pappu", "Batu El", "Hancheng Cao", "Carmelo di Nolfo", "Yanchao Sun", "Meng Cao", "James Zou" ], "version": "v4", "submission_year": 2026, "venue_year": 2026, "publication_status": "accepted", "venue": "ICML", "track": "Regular", "official_url": "https://arxiv.org/abs/2602.01011v4", "venue_url": "https://openreview.net/forum?id=Xn8kmKvO9g", "accessed_at": "2026-07-15" }, { "source_id": "S22", "source_kind": "arxiv", "paper_id": "2604.02460", "title": "Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets", "authors": [ "Dat Tran", "Douwe Kiela" ], "version": "v2", "submission_year": 2026, "venue_year": null, "publication_status": "preprint", "venue": null, "track": null, "official_url": "https://arxiv.org/abs/2604.02460v2", "venue_url": null, "accessed_at": "2026-07-15" }, { "source_id": "S23", "source_kind": "arxiv", "paper_id": "2601.12307", "title": "Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline", "authors": [ "Jiawei Xu", "Arief Koesdwiady", "Sisong Bei", "Yan Han", "Baixiang Huang", "Dakuo Wang", "Yutong Chen", "Zheshen Wang", "Peihao Wang", "Pan Li", "Ying Ding" ], "version": "v1", "submission_year": 2026, "venue_year": null, "publication_status": "preprint", "venue": null, "track": null, "official_url": "https://arxiv.org/abs/2601.12307v1", "venue_url": null, "accessed_at": "2026-07-15" }, { "source_id": "S24", "source_kind": "arxiv", "paper_id": "2604.07821", "title": "More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration", "authors": [ "Advait Yadav", "Sid Black", "Oliver Sourbut" ], "version": "v2", "submission_year": 2026, "venue_year": 2026, "publication_status": "published", "venue": "ICML", "track": "Main Conference Poster", "official_url": "https://arxiv.org/abs/2604.07821v2", "venue_url": "https://icml.cc/virtual/2026/poster/60576", "accessed_at": "2026-07-15" }, { "source_id": "S25", "source_kind": "arxiv", "paper_id": "2603.01045", "title": "Silo-Bench: A Scalable Environment for Evaluating Distributed Coordination in Multi-Agent LLM Systems", "authors": [ "Yuzhe Zhang", "Feiran Liu", "Yi Shan", "Xinyi Huang", "Xin Yang", "Yueqi Zhu", "Xuxin Cheng", "Cao Liu", "Ke Zeng", "Terry Jingchen Zhang", "Wenyuan Jiang" ], "version": "v2", "submission_year": 2026, "venue_year": 2026, "publication_status": "published", "venue": "ACL", "track": "Long Paper", "official_url": "https://arxiv.org/abs/2603.01045v2", "venue_url": "https://aclanthology.org/2026.acl-long.1354/", "accessed_at": "2026-07-15" }, { "source_id": "S26", "source_kind": "arxiv", "paper_id": "2505.11556", "title": "Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs", "authors": [ "Yuxuan Li", "Aoi Naito", "Hirokazu Shirado" ], "version": "v4", "submission_year": 2025, "venue_year": 2026, "publication_status": "published", "venue": "ICML", "track": "Main Conference Poster", "official_url": "https://arxiv.org/abs/2505.11556v4", "venue_url": "https://icml.cc/virtual/2026/papers.html", "accessed_at": "2026-07-15" }, { "source_id": "S27", "source_kind": "openreview", "paper_id": null, "external_id": "manXhfpRH3", "title": "On the Role of Learned Alignment Matrices in LatentMAS", "authors": [ "Spursh Deshpande", "Wenhao Lu" ], "version": "OpenReview revision 2026-06-20", "submission_year": 2026, "venue_year": 2026, "publication_status": "workshop", "venue": "ICML", "track": "CompLearn Workshop Poster", "official_url": "https://openreview.net/forum?id=manXhfpRH3", "venue_url": "https://openreview.net/forum?id=manXhfpRH3", "accessed_at": "2026-07-15" } ], "papers": [ { "paper_id": "2303.17760", "source_id": "S01", "group": "foundation", "research_question": "角色提示能否让两个 LLM 在低人工干预下持续协作？", "mechanism": "任务细化 + AI 用户/AI 助手双角色 + inception prompting + 终止规则。", "experiment": "GPT-3.5；AI Society 与 Code 各 100 个任务；453 个人类偏好。", "result": "AI Society 人类偏好 CAMEL 76.3%，one-shot 10.4%，平局 13.3%；GPT-4 judge 在两组给出 73%/76% 胜率。", "limitations": "主要是主观偏好；仅双 agent、最多 40 条消息；judge 有长度偏好。", "selection": { "decision": "included", "reasons": [ "角色协作早期代表", "正式 venue", "有直接对照" ] } }, { "paper_id": "2305.14325", "source_id": "S02", "group": "foundation", "research_question": "同模型多副本互看并修订答案能否提高推理与事实性？", "mechanism": "通常 3 个独立回答者、2 轮互相检查与修订。", "experiment": "算术、GSM、棋类、人物传记、MMLU；与单 agent、反思、投票对照。", "result": "单 agent 到 debate：算术 67.0→81.8、GSM 77.0→85.0、MMLU 63.9→71.1。", "limitations": "会形成高置信错误共识；长对话更关注最近消息；计算成本更高。", "selection": { "decision": "included", "reasons": [ "辩论范式代表", "多任务直接对照", "正式 venue" ] } }, { "paper_id": "2307.07924", "source_id": "S03", "group": "foundation", "research_question": "角色化语言通信能否覆盖软件设计、编码与测试流水线？", "mechanism": "Waterfall chat chain、阶段性双角色、主动澄清与短记忆交接。", "experiment": "1,200 个 SRDD 需求；GPT-3.5；最多 10 轮；对照 GPT-Engineer 与 MetaGPT。", "result": "可执行率 .8800，对照 .3583/.4145；去角色后可执行率降至 .58；代价为 22,949 tokens、148.2 秒。", "limitations": "产物偏简单原型；未充分测功能正确性、安全、UX；成本显著增加。", "selection": { "decision": "included", "reasons": [ "完整角色流水线", "大样本任务集", "有消融与成本" ] } }, { "paper_id": "2308.00352", "source_id": "S04", "group": "foundation", "research_question": "SOP、专业角色与结构化中间产物能否减少软件团队级联错误？", "mechanism": "PM/架构/工程/QA 角色按 SOP 串行工作，以 PRD、接口、代码等工件交接。", "experiment": "HumanEval、MBPP 与 70 个软件需求；软件人评抽取 7 个，角色消融 2 个。", "result": "HumanEval/MBPP pass@1 85.9%/87.7%；执行反馈带来 +4.2/+5.4 点。", "limitations": "端到端软件人评与角色消融样本极小，且需求由系统生成。", "selection": { "decision": "included", "reasons": [ "SOP 与工件协作代表", "正式 venue", "含消融" ] } }, { "paper_id": "2308.08155", "source_id": "S05", "group": "foundation", "research_question": "如何把 LLM、人类、工具和代码执行组合成可复用多 agent 应用？", "mechanism": "Conversable Agent 统一消息与回复，支持函数/代码、人类输入和静态或动态 GroupChat。", "experiment": "MATH、RAG、ALFWorld、安全 OptiGuide、动态群聊、国际象棋六类应用。", "result": "MATH level-5 成功率 52.5%；ALFWorld 加 grounding agent 后 54%→69%；安全编码 F1 提高 8/35 点。", "limitations": "各应用同时改变模型、提示、工具和拓扑，无法把全部收益归因于多 agent。", "selection": { "decision": "included", "reasons": [ "编排基础设施代表", "多类应用", "正式 venue" ] } }, { "paper_id": "2309.13007", "source_id": "S06", "group": "foundation", "research_question": "异构模型与置信度加权能否改善同模型辩论的相关错误？", "mechanism": "不同 LLM 圆桌讨论、多轮说服、置信度加权投票。", "experiment": "多类常识与推理 benchmark；同单模型、投票和讨论基线比较。", "result": "跨任务结果支持模型多样性与多轮讨论的互补收益，但并非每题都能消除错误共识。", "limitations": "商业模型版本与成本会漂移；置信度不天然校准；仍依赖语言 judge。", "selection": { "decision": "included", "reasons": [ "异构辩论代表", "处理错误相关性", "正式 venue" ] } }, { "paper_id": "2310.02170", "source_id": "S07", "group": "foundation", "research_question": "能否按任务自动选人、裁剪低价值参与者并早停？", "mechanism": "Team Optimization 学习 agent 重要性；任务时动态网络、逐轮排序和 2/3 一致早停。", "experiment": "HumanEval、WebShop、MMLU、MATH；主要 GPT-3.5；与单 agent、辩论等比较。", "result": "HumanEval 73.2→82.9；MMLU 66.4→70.5 且平均 4.39 calls；优化后调用 23.04→16.85。", "limitations": "报告调用数未充分计一次性团队优化成本；只有重复任务才易摊销。", "selection": { "decision": "included", "reasons": [ "动态组队代表", "同时报告质量与调用", "正式 venue" ] } }, { "paper_id": "2402.05120", "source_id": "S08", "group": "foundation", "research_question": "不通信只增加独立采样与投票，能否得到所谓多 agent 增益？", "mechanism": "同一输入独立采样 n 次，以多数票或相似度聚合；不是持续协作。", "experiment": "GSM8K、MATH、Chess、MMLU、HumanEval；最多 40 个实例。", "result": "GPT-3.5 n=1→40：GSM8K .73→.85、MATH .29→.39、MMLU .59→.70。", "limitations": "token 近似线性增长；极难任务增益趋零；不证明角色与通信有效。", "selection": { "decision": "included", "reasons": [ "复杂团队必备简单基线", "多任务规模曲线", "正式期刊" ] } }, { "paper_id": "2406.07155", "source_id": "S09", "group": null, "research_question": "通信拓扑和 agent 数量如何共同影响大规模协作？", "mechanism": "将 agent 组织为有向无环图，并系统改变拓扑和节点规模。", "experiment": "推理与代码任务；多种 topology；展示千级节点运行。", "result": "证明拓扑与规模是独立设计变量，但增益并非可脱离任务、预算和相关错误解释。", "limitations": "大规模运行不等于等成本优势；后续研究发现同质规模会饱和或负增益。", "selection": { "decision": "extended", "reasons": [ "大规模拓扑仍保留为扩展证据", "由 matched-budget 与专家稀释直接边界替换核心席位" ] } }, { "paper_id": "2502.04180", "source_id": "S10", "group": "frontier", "research_question": "能否针对每个查询自动搜索 agent 工作流并兼顾成本？", "mechanism": "把 agent、操作和连接组成概率 supernet，学习查询条件化架构。", "experiment": "多类推理与代码 benchmark；与手工 MAS 与搜索基线比较。", "result": "论文报告相对基线 +0.54 至 +11.82 的性能增益，并降低约 6%–45% 推理成本。", "limitations": "搜索/训练成本需摊销；收益依赖候选组件、代理指标和任务分布稳定性。", "selection": { "decision": "included", "reasons": [ "自动架构搜索前沿", "质量成本联合目标", "ICML 2025" ] } }, { "paper_id": "2502.11133", "source_id": "S11", "group": null, "research_question": "能否同时路由协作模式、角色和底层模型？", "mechanism": "级联控制器按查询依次选择 MAS 模式、角色配置与模型。", "experiment": "数学与代码任务；跨多模型与协作模式比较。", "result": "论文报告 MBPP 提升 1.8–8.2 点，HumanEval 开销最高减少 52.07%。", "limitations": "仍是预印本；路由决策依赖训练分布，未充分覆盖线上漂移与失败回退。", "selection": { "decision": "extended", "reasons": [ "路由方向有价值", "预印本且与 MaAS/DyLAN 机制重叠" ] } }, { "paper_id": "2503.01935", "source_id": "S12", "group": "frontier", "research_question": "如何同时评估任务完成、规划、通信和竞争协作？", "mechanism": "MARBLE 支持 star/tree/chain/graph 与多种规划策略，以动态 milestones 评估。", "experiment": "研究、Minecraft、数据库、编码、谈判、狼人杀六场景；任务场景各 100 cases。", "result": "1→3 agents 后继续增加使 KPI 下降；group discussion 在该实验各指标最差；协调高不保证任务成功。", "limitations": "多项指标依赖 LLM judge；拓扑结论主要来自研究场景，外推有限。", "selection": { "decision": "included", "reasons": [ "过程型评测代表", "含拓扑/规模消融", "ACL 2025" ] } }, { "paper_id": "2503.13657", "source_id": "S13", "group": "frontier", "research_question": "主流多 agent 系统为何失败，能否形成可操作 taxonomy？", "mechanism": "6 位专家用 grounded theory 建立 14 种 failure modes，再用校准 judge 扩展标注。", "experiment": "1,642 条 traces、7 个框架；人类 κ=.88，judge accuracy .94、F1 .80。", "result": "失败率 41%–86.7%；重复、错误终止、规格违背、reasoning-action mismatch 和错误验证突出；同模型结构干预最高 +15.6%。", "limitations": "完整数据多数由 LLM judge 标注；taxonomy 不声称穷尽，相关症状不等于因果。", "selection": { "decision": "included", "reasons": [ "真实轨迹失败诊断", "人工校准", "NeurIPS 数据与基准 track" ] } }, { "paper_id": "2505.21471", "source_id": "S14", "group": null, "research_question": "多 agent 能否分布式处理远超单模型上下文的外部知识？", "mechanism": "将外部知识切分给多个 agent，通过同步、协调和聚合回答。", "experiment": "长上下文/外部知识任务，对不同分片与协调设置比较。", "result": "支持通过分布式上下文扩展可处理证据量，但质量依赖分片可分解性与聚合完整性。", "limitations": "切分会丢跨分片关系；同步与聚合开销高；不等同于安全、持久的共享记忆。", "selection": { "decision": "extended", "reasons": [ "长上下文分片机制保留", "由更直接的 matched-budget 反证替换核心席位" ] } }, { "paper_id": "2511.20639", "source_id": "S15", "group": "frontier", "research_question": "agent 是否必须用自然语言通信，还是可共享 latent working memory？", "mechanism": "在连续 hidden representation 中传递信息，减少文本化与重复编码。", "experiment": "9 个 benchmarks；多种协作拓扑与模型；与文本通信比较。", "result": "原论文在 9 个 benchmark 报告最高 +14.6 准确率、输出 token 减少 70.8%–83.7%、速度约 4–4.3 倍；后续直接复核显示这些收益并不稳定依赖 learned alignment matrix。", "limitations": "CompLearn 2026 复核中，identity matrix 无显著下降；部分任务移除跨 agent transfer 相当或更好，代码任务文本通信优于 latent。latent 状态的审计、互操作与安全问题仍未解决。", "selection": { "decision": "included", "reasons": [ "latent communication 前沿", "多 benchmark", "ICML Spotlight" ] } }, { "paper_id": "2512.08296", "source_id": "S16", "group": "frontier", "research_question": "在控制工具、提示和计算后，何时多 agent 扩展有效或有害？", "mechanism": "统一比较单 agent、独立、中心化、去中心化与混合架构。", "experiment": "260 个配置、6 个 agentic benchmarks、3 个 LLM family。", "result": "相对单 agent 从可分解财务推理 +80.8% 到序列规划 -70.0%；模型为留出配置选对架构 87%。", "limitations": "预印本；跨 benchmark 模型解释度有限；数字不能直接迁移到不同工具与预算。", "selection": { "decision": "included", "reasons": [ "等预算受控规模研究", "同时呈现正负效应", "最新架构-任务匹配证据" ] } }, { "paper_id": "2602.03794", "source_id": "S17", "group": null, "research_question": "为什么同质 agent 扩展饱和，而多样性可能继续带来收益？", "mechanism": "以有效信息通道 K* 描述输出相关性与互补证据。", "experiment": "不同模型、提示、工具的同质/异质配置与规模曲线。", "result": "论文报告 2 个多样化 agent 可匹配或超过 16 个同质 agent。", "limitations": "workshop 论文；K* 与实际产品任务的在线可估计性仍需验证。", "selection": { "decision": "extended", "reasons": [ "多样性机制保留为扩展证据", "workshop 证据弱于 2026 顶会直接协调失败研究" ] } }, { "paper_id": "2602.01566", "source_id": "S18", "group": null, "research_question": "文件系统能否作为长周期研究的外部记忆与跨 agent 协调介质？", "mechanism": "Context Builder 构建层级知识库；Report Writer 分段读取并写作。", "experiment": "DeepResearch Bench 与 DeepConsult；不同 backbone 和 Context Builder 计算预算。", "result": "在两个开放研究 benchmark 报告 SOTA，且报告质量与 Context Builder 计算量正相关。", "limitations": "两 agent、研究写作单域；文件系统的权限、污染、删除与冲突治理未被充分评价。", "selection": { "decision": "extended", "reasons": [ "深度研究编排保留为扩展证据", "主题专门性弱于 2026 顶会直接协调失败研究" ] } }, { "paper_id": "2308.10848", "source_id": "S19", "group": null, "selection": { "decision": "extended", "reasons": [ "与 AutoGen/DyLAN 重叠", "保留错误同伴说服正确 agent 的正式反证" ] } }, { "paper_id": "2606.18829", "source_id": "S20", "group": null, "selection": { "decision": "extended", "reasons": [ "共享记忆治理直接相关", "截止日时仍是单篇新预印本" ] } }, { "paper_id": "2602.01011", "source_id": "S21", "group": "frontier", "research_question": "团队已知谁是专家时，讨论能否保留最佳个体的专业判断？", "mechanism": "对照无专家、隐含专家、显式揭示专家及最佳个体；改变团队规模与专长分布。", "experiment": "四类协作/排序任务，多模型与团队构成；比较团队与已识别最佳个体。", "result": "即使明确揭示专家，团队仍未能匹配最佳个体，差距为 6.3%–41.1%；规模增加会加剧判断稀释。", "limitations": "受控任务不等同于生产组织；共识有时改善鲁棒性，但会牺牲最佳专家信号。", "selection": { "decision": "included", "reasons": [ "ICML 2026", "直接揭示专家稀释", "重审共识与规模结论所需反证" ] } }, { "paper_id": "2604.02460", "source_id": "S22", "group": "frontier", "research_question": "在相同 thinking-token 预算下，MAS 是否仍优于单智能体？", "mechanism": "跨三个模型家族配平总思考 token，对比 single-agent scaling 与多种 MAS。", "experiment": "两个 multi-hop reasoning 数据集；多预算点、模型家族和信息退化条件。", "result": "大多数配平条件下单智能体匹配或超过 MAS；当单体上下文被人为退化时 MAS 才显示边界性优势。", "limitations": "集中于 multi-hop reasoning；不能替代持续行动、异构权限或外部状态任务评测。", "selection": { "decision": "included", "reasons": [ "直接 matched-budget 证据", "约束 C05 强度", "覆盖模型家族与预算曲线" ] } }, { "paper_id": "2601.12307", "source_id": "S23", "group": "frontier", "research_question": "固定同一模型的多智能体 workflow 能否折叠成单智能体执行？", "mechanism": "OneFlow 由单智能体顺序执行原 workflow 节点，并复用 KV cache。", "experiment": "多 workflow、模型和任务；比较效果、token/延迟及异构模型边界。", "result": "同质 workflow 可匹配或略优于 MAS 并节省 KV-cache/通信开销；真正异构模型组合不保证可折叠。", "limitations": "结论针对同质 workflow；不同权限、并行环境交互和独立信息源仍可能需要团队。", "selection": { "decision": "included", "reasons": [ "直接 strong single-agent baseline", "揭示同质 workflow 可折叠", "补足复杂 MAS 对照契约" ] } }, { "paper_id": "2604.07821", "source_id": "S24", "group": "frontier", "research_question": "模型能力提升是否自然带来无需额外激励的合作能力？", "mechanism": "在零成本协作博弈中分离个体能力与合作表现，并测试协议和微激励干预。", "experiment": "多类 LLM、最优集体基线，以及结构化协议和微激励条件。", "result": "能力不能预测合作：o3 仅达到最优集体表现的 17%；结构化协议可使部分低能力模型翻倍，微激励可改善弱合作模型。", "limitations": "零成本博弈不覆盖长期工具工作流；干预效果不能外推为通用合作解法。", "selection": { "decision": "included", "reasons": [ "ICML 2026 主会", "直接分离能力与合作", "包含协议与激励干预" ] } }, { "paper_id": "2603.01045", "source_id": "S25", "group": "frontier", "research_question": "agent 能否把分布式私有信息转化为可靠的联合推理？", "mechanism": "Silo-Bench 将信息分布到 agent silo，并区分通信是否到位与推理是否正确。", "experiment": "30 个任务；6 种团队规模 × 3 种通信协议 × 3 个模型 = 54 个配置，共 1,620 次实验。", "result": "Communication–Reasoning Gap 显示消息到达不等于信息被正确整合；三类主要失败为 Premature Submission（过早提交）37.2%、Consensus Failure（共识失败：多个 agent 提交不同答案且未同步）29.9% 和 Computation Error（计算错误）28.6%。", "limitations": "基准任务仍是受控环境；生产系统的权限、延迟和异步失败更复杂。", "selection": { "decision": "included", "reasons": [ "ACL 2026 Main/Long", "直接评测分布式协调", "大规模配置与失败分解" ] } }, { "paper_id": "2505.11556", "source_id": "S26", "group": null, "research_question": "分布式信息条件下，多 agent 能否完成可靠的集体决策推理？", "mechanism": "HiddenBench 将完成任务所需信息分散到不同 agent，并与获得完整信息的单 agent 对照。", "experiment": "65 个 hidden-profile 任务、15 个模型；比较分布式 MAS 与拥有完整信息的单 agent。", "result": "分布式 MAS 准确率 30.1%，完整信息单 agent 准确率 80.7%；扩大模型并未消除信息整合失败。", "limitations": "65 个受控 hidden-profile 任务不能代表全部长期协作；与 Silo-Bench 的任务结构不同。", "selection": { "decision": "extended", "reasons": [ "ICML 2026，直接研究分布式信息整合失败", "与 Silo-Bench 形成前沿交叉验证", "与核心 Silo-Bench 边界高度重叠，保留 18 篇核心上限" ] } }, { "paper_id": "2309.17288", "title": "AutoAgents: A Framework for Automatic Agent Generation", "version": "v3", "submission_year": 2023, "publication_status": "published", "source_url": "https://arxiv.org/abs/2309.17288v3", "selection": { "decision": "extended", "reasons": [ "自动组队前驱；证据弱于 DyLAN/MaAS" ] } }, { "paper_id": "2402.14034", "title": "AgentScope: A Flexible yet Robust Multi-Agent Platform", "version": "v2", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2402.14034v2", "selection": { "decision": "extended", "reasons": [ "平台工程贡献为主，机制因果较弱" ] } }, { "paper_id": "2402.16823", "title": "GPTSwarm: Language Agents as Optimizable Graphs", "version": "v3", "submission_year": 2024, "publication_status": "published", "source_url": "https://arxiv.org/abs/2402.16823v3", "selection": { "decision": "extended", "reasons": [ "重要前驱，MaAS 覆盖更直接的架构搜索" ] } }, { "paper_id": "2406.04692", "title": "Mixture-of-Agents Enhances Large Language Model Capabilities", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2406.04692v1", "selection": { "decision": "extended", "reasons": [ "token 聚合，不是持续行动团队" ] } }, { "paper_id": "2406.05720", "title": "VillagerAgent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in Minecraft", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2406.05720v1", "selection": { "decision": "extended", "reasons": [ "Minecraft 单域色彩强" ] } }, { "paper_id": "2410.02506", "title": "Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2410.02506v1", "selection": { "decision": "extended", "reasons": [ "通信剪枝有价值，但核心优先跨系统边界" ] } }, { "paper_id": "2411.04468", "title": "Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2411.04468v1", "selection": { "decision": "extended", "reasons": [ "工程基线强，机制隔离与正式 venue 较弱" ] } }, { "paper_id": "2412.05449", "title": "Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2412.05449v1", "selection": { "decision": "extended", "reasons": [ "专有系统、手工场景和对照有限" ] } }, { "paper_id": "2505.18279", "title": "Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control", "version": "v1", "submission_year": 2025, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2505.18279v1", "selection": { "decision": "extended", "reasons": [ "provenance/ACL 有启发，尚无任务实证" ] } }, { "paper_id": "2604.19278", "title": "Explicit Trait Inference for Multi-Agent Coordination", "version": "v2", "submission_year": 2026, "publication_status": "accepted", "source_url": "https://arxiv.org/abs/2604.19278v2", "selection": { "decision": "extended", "reasons": [ "ACL 2026 新机制，与协调失败路径部分重叠" ] } }, { "paper_id": "2502.19559", "title": "Stay Focused: Problem Drift in Multi-Agent Debate", "version": "v3", "submission_year": 2025, "publication_status": "accepted", "source_url": "https://arxiv.org/abs/2502.19559v3", "selection": { "decision": "extended", "reasons": [ "正式 debate 反证，保留为扩展证据候选" ] } }, { "paper_id": "2604.13349", "title": "When Less Latent Leads to Better Relay: Information-Preserving Compression for Latent Multi-Agent LLM Collaboration", "version": "v2", "submission_year": 2026, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2604.13349v2", "selection": { "decision": "extended", "reasons": [ "LatentMAS 后续，非常新" ] } }, { "paper_id": "2402.01680", "title": "Large Language Model based Multi-Agents: A Survey of Progress and Challenges", "version": "v2", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2402.01680v2", "selection": { "decision": "extended", "reasons": [ "综述用于召回，不作一手机制证据" ] } }, { "paper_id": "2501.06322", "title": "Multi-Agent Collaboration Mechanisms: A Survey of LLMs", "version": "v1", "submission_year": 2025, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2501.06322v1", "selection": { "decision": "extended", "reasons": [ "分类学参考，非原始实验" ] } }, { "paper_id": "2502.14321", "title": "Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems", "version": "v3", "submission_year": 2025, "publication_status": "accepted", "source_url": "https://arxiv.org/abs/2502.14321v3", "selection": { "decision": "extended", "reasons": [ "通信路径索引，非一手因果证据" ] } }, { "paper_id": "2505.07313", "title": "Towards Multi-Agent Reasoning Systems for Collaborative Expertise Delegation: An Exploratory Design Study", "version": "v2", "submission_year": 2025, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2505.07313v2", "selection": { "decision": "extended", "reasons": [ "探索性 preprint，验证范围较弱" ] } }, { "paper_id": "2510.13821", "title": "LLM Agent Communication Protocol (LACP) Requires Urgent Standardization: A Telecom-Inspired Protocol is Necessary", "version": "v1", "submission_year": 2025, "publication_status": "workshop", "source_url": "https://arxiv.org/abs/2510.13821v1", "selection": { "decision": "extended", "reasons": [ "协议 position paper 为主，实证不足" ] } }, { "paper_id": "2506.03053", "title": "MAEBE: Multi-Agent Emergent Behavior Framework", "version": "v2", "submission_year": 2025, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2506.03053v2", "selection": { "decision": "extended", "reasons": [ "同伴压力方向重要，验证与 venue 较弱" ] } }, { "paper_id": "2502.20073", "title": "Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents", "version": "v3", "submission_year": 2025, "publication_status": "published", "source_url": "https://arxiv.org/abs/2502.20073v3", "selection": { "decision": "extended", "reasons": [ "EMNLP 2025；交互长链评测，扩展保留" ] } }, { "paper_id": "2408.00989", "title": "On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents", "version": "v4", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2408.00989v4", "selection": { "decision": "extended", "reasons": [ "故障传播与结构韧性，扩展反证" ] } }, { "paper_id": "2410.07283", "title": "Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems", "version": "v1", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2410.07283v1", "selection": { "decision": "extended", "reasons": [ "多 agent 攻击面，安全专题候选" ] } }, { "paper_id": "2407.04622", "title": "On scalable oversight with weak LLMs judging strong LLMs", "version": "v2", "submission_year": 2024, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2407.04622v2", "selection": { "decision": "extended", "reasons": [ "debate 的任务依赖边界，扩展反证" ] } }, { "paper_id": "2505.23352", "title": "Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems", "version": "v1", "submission_year": 2025, "publication_status": "preprint", "source_url": "https://arxiv.org/abs/2505.23352v1", "selection": { "decision": "extended", "reasons": [ "拓扑抑错与信息扩散双刃效应，扩展反证" ] } } ], "evidence": [ { "evidence_id": "E01", "source_id": "S01", "locator": "§5.1 Table 1; Appendix K.1", "kind": "paper_result", "faithful_summary": "双角色 CAMEL 在 453 个人类偏好中获 76.3% 偏好，one-shot 10.4%，但指标是偏好而非事实正确率。", "verified_by": [ "Reader", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E02", "source_id": "S02", "locator": "§3.1 Table 1; §3.2 Table 2; §5", "kind": "paper_result", "faithful_summary": "3-agent、2-round debate 在六类任务优于匹配单 agent；作者同时报告错误共识、高置信错误和长上下文近因偏好。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E03", "source_id": "S03", "locator": "§4 Tables 1–4; §6", "kind": "paper_result", "faithful_summary": "ChatDev 可执行率 .8800，去角色后 .58；相较单体基线 token 和延迟显著上升，产物多为简单原型。", "verified_by": [ "Reader" ], "verified_at": "2026-07-15" }, { "evidence_id": "E04", "source_id": "S04", "locator": "§4.1–4.4 Tables 1–3; Appendix D", "kind": "paper_result", "faithful_summary": "MetaGPT 的 SOP、结构化工件和执行反馈有正向结果；端到端软件人评仅 7 个任务、角色消融 2 个。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E05", "source_id": "S05", "locator": "§2.1–2.2; §3 A1–A6; Figure 3–4; §4", "kind": "paper_result", "faithful_summary": "AutoGen 六类应用展示可编程会话编排；各应用同时改变多个因素，不能作为多 agent 普遍优越的因果证据。", "verified_by": [ "Reader" ], "verified_at": "2026-07-15" }, { "evidence_id": "E06", "source_id": "S06", "locator": "§3–§5 and official abstract", "kind": "paper_result", "faithful_summary": "异构 LLM、多轮讨论与置信度加权在多项推理任务改善聚合，但置信度与共识仍不等于外部验证。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E07", "source_id": "S07", "locator": "§3.1, §3.3–3.4; §4 Tables 2–6", "kind": "paper_result", "faithful_summary": "动态选人和早停可提高任务阶段质量并减少调用；一次性 Team Optimization 成本未完整计入。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E08", "source_id": "S08", "locator": "§4 Figure 3 Table 2; §5.5; §6–7", "kind": "paper_result", "faithful_summary": "无通信独立采样和投票已在多个任务随 n 提升，但 token 近线性增长且困难任务趋于饱和。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E09", "source_id": "S09", "locator": "Abstract; §3–§5", "kind": "paper_result", "faithful_summary": "MacNet 证明拓扑和节点规模可系统操纵并可运行到千级，但不提供跨任务等预算的普遍规模定律。", "verified_by": [ "Scout", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E10", "source_id": "S10", "locator": "Abstract; Tables 1–3", "kind": "paper_result", "faithful_summary": "MaAS 报告查询条件化架构搜索带来 0.54–11.82 性能增益并降低约 6%–45% 推理成本；搜索成本需另计。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E11", "source_id": "S11", "locator": "Abstract; §4", "kind": "paper_result", "faithful_summary": "MasRouter 联合选择协作模式、角色和模型，报告代码任务增益与最高 52.07% 开销下降；结果仍为预印本。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E12", "source_id": "S12", "locator": "§4 Table 1 Figures 5–6; §5 Figures 7–8", "kind": "paper_result", "faithful_summary": "MultiAgentBench 中高协调分不保证高任务分，group discussion 表现差，agent 从 3 继续增加时 KPI 下降。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E13", "source_id": "S13", "locator": "§3 Tables 1–2; §4 Figure 1; §5; Appendix H/J", "kind": "paper_result", "faithful_summary": "1,642 条轨迹揭示 14 类故障；重复、错误终止、规格违背、行动错配和错误验证是独立于最终答案的故障源。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E14", "source_id": "S13", "locator": "§6.2 Table 4", "kind": "paper_result", "faithful_summary": "MAST 对 ChatDev 的 improved prompt 同时调整角色层级、仅允许上级结束对话，并要求 verifier 关注 edge cases；ProgramDev 由 25.0% 升至 34.4%（+9.4pp）。这是组合式 role/prompt intervention，不能归因于终止权单项。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E15", "source_id": "S14", "locator": "Abstract; §3–§5", "kind": "paper_result", "faithful_summary": "ExtAgents 通过分布式上下文处理超窗口证据；收益依赖可分片性、同步与完整聚合。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E16", "source_id": "S15", "locator": "§4.1 Table 1; Appendix D Tables 3–5", "kind": "paper_result", "faithful_summary": "LatentMAS 在 9 个 benchmark 报告最高 +14.6 准确率、70.8%–83.7% 输出 token 降低及约 4 倍速度，但 latent 状态降低审计性。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E17", "source_id": "S16", "locator": "§4.2–§4.4; Figure 1; Table 2", "kind": "paper_result", "faithful_summary": "260 配置受控实验中，多 agent 相对单 agent 从 +80.8% 到 -70.0%，架构-任务对齐决定方向。", "verified_by": [ "Scout", "PM-Paper", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E18", "source_id": "S17", "locator": "§4.1–§4.3; Figures 2–4; Tables 1–2", "kind": "paper_result", "faithful_summary": "同质 agent 输出强相关而快速饱和；论文报告 2 个多样化 agent 可匹配或超过 16 个同质 agent。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E19", "source_id": "S18", "locator": "§2.1–§2.3; §3.2 Tables 2–4; §4.1 Figure 4", "kind": "paper_result", "faithful_summary": "文件系统作为跨 agent、跨会话持久工件与协调介质；两个开放研究 benchmark 上报告 SOTA。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E20", "source_id": "S19", "locator": "§3 Tables 1–2; §4 Figure 6; Appendix F", "kind": "contrary_result", "faithful_summary": "GPT-3.5 Group 在 CommonGen/MGSM 低于 Solo，约 10% MGSM 错误来自正确 agent 被错误反馈说服。", "verified_by": [ "Reader", "Critic" ], "verified_at": "2026-07-15" }, { "evidence_id": "E21", "source_id": "S20", "locator": "§3–§5; Tables 1–3; Appendix B", "kind": "emerging_result", "faithful_summary": "GateMem 联合评估长期效用、情境授权边界与主动遗忘；不同基线尚不能同时实现高效用、稳健访问控制和可靠删除，因此截至 cutoff 仅作新兴设计信号。", "verified_by": [ "Scout", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E22", "source_id": "S21", "locator": "§4.1 Table 1; §4.2 Table 2; §4.3; Appendix H", "kind": "contrary_result", "faithful_summary": "团队即使被明确告知谁是专家，仍未匹配最佳个体；四类条件差距为 6.3%–41.1%，更多成员会通过共识压力稀释专家判断。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E23", "source_id": "S22", "locator": "§4.4 Table 1; §5.1; §5.3 Figure 3; Appendix F Tables 3–10", "kind": "contrary_result", "faithful_summary": "在三个模型家族、两个 multi-hop 数据集和相同 thinking-token 预算下，single-agent scaling 多数匹配或超过 MAS；MAS 优势主要出现在单体上下文被退化的边界条件。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E24", "source_id": "S23", "locator": "§4.2 Tables 1–2; §4.2.3 Table 3; §4.2.4 Table 4", "kind": "contrary_result", "faithful_summary": "OneFlow 让单智能体执行同质 MAS workflow，可匹配或略优于原系统并减少 KV-cache/通信开销；异构模型或真实并行交互不保证可折叠。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E25", "source_id": "S24", "locator": "Abstract; §4–§6; Table 3", "kind": "contrary_result", "faithful_summary": "个体能力不预测零成本合作：o3 仅达到最优集体表现的 17%，而 o3-mini 为 50%；显式协议可使部分低能力模型表现翻倍，微激励能改善弱合作模型。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E26", "source_id": "S25", "locator": "§3–§5; §5.2 Table 4; Figures 4–6", "kind": "contrary_result", "faithful_summary": "Silo-Bench 用 30 个任务，在 6 种团队规模 × 3 种通信协议 × 3 个模型形成的 54 个配置上完成 1,620 次实验，识别 Communication–Reasoning Gap；Table 4 的三类主要失败是 Premature Submission（过早提交）37.2%、Consensus Failure（共识失败：多个 agent 提交不同答案且未同步）29.9% 和 Computation Error（计算错误）28.6%。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E27", "source_id": "S26", "locator": "Abstract; §3; §4.2; Tables 1–2; §5", "kind": "contrary_result", "faithful_summary": "HiddenBench 在 65 个 hidden-profile 任务和 15 个模型上的准确率结果显示：分布式 MAS 为 30.1%，而获得完整信息的单 agent 为 80.7%；信息不对称下的整合失败不会随模型规模可靠消失。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" }, { "evidence_id": "E28", "source_id": "S27", "locator": "§4.1 Table 1; §4.2–§4.3; Appendix C", "kind": "contrary_result", "faithful_summary": "对 LatentMAS 的直接复核发现 identity matrix 不造成显著下降；部分任务移除跨 agent transfer 相当或更好，代码任务中文本通信优于 latent communication。", "verified_by": [ "Reader", "Critic", "PM-Paper" ], "verified_at": "2026-07-15" } ], "claims": [ { "claim_id": "C01", "type": "synthesis", "text": "多 agent 不是天然优于单 agent；任务可分解性、独立信息、架构匹配、预算与验证方式共同决定收益方向。", "strength": "strong", "scope": "受控任务与现有 benchmark；不等同于所有生产工作流", "supporting_evidence_ids": [ "E12", "E13", "E17", "E23", "E24" ], "contradicting_evidence_ids": [ "E01", "E03", "E10" ], "limitations": "正向案例常同时改变 prompt、工具或预算。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E01" }, { "evidence_id": "E03" }, { "evidence_id": "E10" } ], "no_contrary_note": null, "revision_reason": "检索 single-agent、best-of-n、matched-budget 与正向 MAS；保留‘取决于’而非‘普遍失败’。" } }, { "claim_id": "C02", "type": "synthesis", "text": "角色与 SOP 的可靠价值来自明确接口、结构化工件、权限、终止条件和可执行反馈，而不是角色名称本身；现有 MAST 证据是组合式 role/prompt intervention，未隔离终止权单项因果。", "strength": "conditional", "scope": "证据最集中于阶段化的软件工程和双角色任务", "supporting_evidence_ids": [ "E03", "E04", "E14" ], "contradicting_evidence_ids": [ "E24" ], "limitations": "开放任务、真实组织和长期演进证据不足；+9.4pp 不能归因于单个终止权变更。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E24" } ], "no_contrary_note": null, "revision_reason": "检索 role/SOP ablation 与 single-agent workflow replay；同质 workflow 可折叠，因此限定为接口与工件价值。" } }, { "claim_id": "C03", "type": "synthesis", "text": "同质 agent 的数量或 workflow 容易饱和、折叠甚至稀释已知专家；即使消息成功传递，分布式信息也可能无法被正确整合。真实互补信息、合作协议和可验证专长比席位数更关键。", "strength": "conditional", "scope": "推理与聚合任务；多样性需要真实互补而非 persona 文案差异", "supporting_evidence_ids": [ "E08", "E18", "E22", "E24", "E25", "E26", "E27" ], "contradicting_evidence_ids": [ "E09", "E17" ], "limitations": "专家授权、异构性质量和独立信息量仍缺线上可观测的统一指标。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E09" }, { "evidence_id": "E17" } ], "no_contrary_note": null, "revision_reason": "检索 agent count、diversity、expert reveal、zero-cost cooperation、distributed information 与 topology scaling；保留可运行/部分正收益，否定单调规模律。" } }, { "claim_id": "C04", "type": "synthesis", "text": "现有正收益主要集中于候选差异可交叉检查的任务；共识、置信度和更长讨论不是正确性证明，还可能稀释已识别专家或掩盖信息整合失败。", "strength": "conditional", "scope": "数学、事实问答和可交叉检查推理", "supporting_evidence_ids": [ "E02", "E06" ], "contradicting_evidence_ids": [ "E20", "E22", "E23", "E26", "E27" ], "limitations": "现代模型、开放行动任务与恶意参与者上的复验有限。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E20" }, { "evidence_id": "E22" }, { "evidence_id": "E23" }, { "evidence_id": "E26" }, { "evidence_id": "E27" } ], "no_contrary_note": null, "revision_reason": "检索 debate、consensus、persuasion、expert dilution、distributed information 与 matched-budget；删除未经必要条件实验支持的‘只在’，改为观察性范围。" } }, { "claim_id": "C05", "type": "synthesis", "text": "复杂 MAS 的增益必须在相同总 thinking-token、工具和采样预算下，对照 best single 与 best-of-n；否则不能归因于协作。", "strength": "conditional", "scope": "答案可聚合或能定义总预算的推理任务；持续行动任务需另加环境交互预算", "supporting_evidence_ids": [ "E08", "E23", "E24" ], "contradicting_evidence_ids": [ "E17" ], "limitations": "现有直接 matched-budget 证据集中于 multi-hop reasoning 与同质 workflow，尚非跨任务定律。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E17" } ], "no_contrary_note": null, "revision_reason": "检索 equal token/call budgets、single scaling、best-of-n 与复杂 MAS；补入直接配平证据后仍因任务范围将 strong 降为 conditional。" } }, { "claim_id": "C06", "type": "synthesis", "text": "动态选人、拓扑、路由和早停可以改善质量成本比，但优化开销只有在重复任务中才容易摊销。", "strength": "conditional", "scope": "候选专长可区分、任务分布相对稳定", "supporting_evidence_ids": [ "E07", "E10", "E11" ], "contradicting_evidence_ids": [ "E24" ], "limitations": "分布漂移会使 router 与搜索策略失效。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "qualified", "findings": [ { "evidence_id": "E24" } ], "no_contrary_note": null, "revision_reason": "检索 routing/search/early-stop 的离线成本与同质 workflow 折叠；要求重复任务可摊销并保留 simple baseline。" } }, { "claim_id": "C07", "type": "synthesis", "text": "结构化状态、最小必要消息和外部可执行验证，通常比全量自由聊天更可控。", "strength": "conditional", "scope": "工具调用、软件工程与长周期研究工作流", "supporting_evidence_ids": [ "E04", "E05", "E12", "E19" ], "contradicting_evidence_ids": [ "E16" ], "limitations": "跨领域统一消融仍少；过度结构化也可能抑制探索。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "qualified", "findings": [ { "evidence_id": "E16" } ], "no_contrary_note": null, "revision_reason": "检索 structured artifact、free chat、latent communication 与 auditability；latent state 可能更高效，因此结论限定为可控性而非绝对性能。" } }, { "claim_id": "C08", "type": "synthesis", "text": "共享记忆不仅是召回问题，也是来源、权限、冲突、保留期、删除与污染治理问题。", "strength": "contested", "scope": "长周期、跨 session、多人可写工作空间", "supporting_evidence_ids": [ "E19", "E21" ], "contradicting_evidence_ids": [], "limitations": "权限化记忆的公开实证很新，尚无成熟跨系统基准。", "status": "qualified", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "no_direct_contrary_found", "findings": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 shared memory、ACL、provenance、deletion 与 memory utility；未发现否定治理必要性的直接反证，因证据新仍维持 contested。" } }, { "claim_id": "C09", "type": "synthesis", "text": "Agent Team 评测必须同时看任务闭环、里程碑/trace、成本、延迟与失败归因；“协调分”或最终文本单项都不够。", "strength": "strong", "scope": "可观测的 agent workflow", "supporting_evidence_ids": [ "E12", "E13", "E26", "E27" ], "contradicting_evidence_ids": [], "limitations": "开放任务仍需人工与 LLM judge，评测本身也可能偏差。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "no_direct_contrary_found", "findings": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 outcome-only、coordination score、trace/cost/failure evaluation；Silo-Bench/HiddenBench 补强过程诊断，未发现支持单指标充分性的直接反证。" } }, { "claim_id": "C10", "type": "synthesis", "text": "多 agent 故障不仅来自底模幻觉；信息流、规格遵从、终止、验证和 reasoning-action 对齐都是独立故障面。", "strength": "strong", "scope": "7 个主流框架的编码、数学与通用 agent 轨迹", "supporting_evidence_ids": [ "E13", "E14", "E25", "E26", "E27" ], "contradicting_evidence_ids": [], "limitations": "taxonomy 非穷尽，自动标签主要表示症状而非根因。", "status": "validated", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "no_direct_contrary_found", "findings": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 hallucination-only、cooperation failure、communication-reasoning gap 与 system-level failure taxonomy；未发现将多 agent 故障完整归为底模幻觉的直接反证。" } }, { "claim_id": "C11", "type": "synthesis", "text": "latent communication、自动架构搜索与权限化共享记忆仍有探索价值，但部分 LatentMAS 任务不依赖 learned alignment 或跨 agent transfer，且代码任务文本通信更好；审计、互操作、安全和线上复验仍未形成共识。", "strength": "contested", "scope": "2025–2026 新兴机制", "supporting_evidence_ids": [ "E10", "E16", "E21" ], "contradicting_evidence_ids": [ "E23", "E24", "E28" ], "limitations": "LatentMAS 的 headline 结果已有直接负向复核；其他机制也多为单篇论文或预印本。", "status": "qualified", "show_in_summary": true, "counter_search": { "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "outcome": "contrary_found", "findings": [ { "evidence_id": "E23" }, { "evidence_id": "E24" }, { "evidence_id": "E28" } ], "no_contrary_note": null, "revision_reason": "检索 latent/search/memory 新机制的 simple baseline、audit 和 matched-budget；直接 LatentMAS 复核否定 learned alignment 的稳定必要性，因此收窄前沿判断。" } }, { "claim_id": "R01", "type": "recommendation", "title": "先证明一个 Agent 不够", "text": "默认跑 single-agent 和等预算 best-of-n；只有任务可分解、需要不同权限/工具或独立证据时才升级团队。", "strength": "strong", "scope": "Multica 任务入口与实验平台", "supporting_evidence_ids": [ "E08", "E17", "E22", "E23", "E24" ], "contradicting_evidence_ids": [], "limitations": "单 agent baseline 也必须使用同等工具和总预算。", "status": "validated", "show_in_summary": false, "evidence_basis": "专家团队可低于最佳个体；配平 thinking tokens 后单体常匹配/超过 MAS；同质 workflow 还可折叠。", "assumptions": "任务能定义可比较的结果指标与预算。", "side_effects": "增加基线运行成本，但能阻止无效组织复杂度。" }, { "claim_id": "R02", "type": "recommendation", "title": "把角色写成契约，不写成人设", "text": "每个角色声明输入、输出工件、工具、权限、专长证据、升级与 stop condition；已知专家可保留最终判断或触发独立复核。", "strength": "conditional", "scope": "可阶段化的 Multica issue 与工作流", "supporting_evidence_ids": [ "E03", "E04", "E14", "E22", "E25" ], "contradicting_evidence_ids": [], "limitations": "探索性任务可能需要临时跨界；协议效果依模型能力和任务而变。", "status": "validated", "show_in_summary": false, "evidence_basis": "角色/SOP、专家稀释和协议干预支持显式契约；MAST 的 +9.4pp 来自组合式 role/prompt intervention，未隔离终止权单项。", "assumptions": "工件和状态转移能被机器或人检查。", "side_effects": "前期建模成本上升；契约过细会降低适应性。" }, { "claim_id": "R03", "type": "recommendation", "title": "在查询时选最小团队", "text": "router 先选最小角色/模型/拓扑与预算；同质 workflow 先尝试单智能体执行，只有验证失败、权限分离或证据冲突时扩容。", "strength": "conditional", "scope": "有重复任务与遥测数据的路由层", "supporting_evidence_ids": [ "E07", "E10", "E11", "E24" ], "contradicting_evidence_ids": [], "limitations": "router 本身需要监控漂移与回退。", "status": "validated", "show_in_summary": false, "evidence_basis": "动态选人、架构搜索和联合路由报告质量成本收益。", "assumptions": "能收集团队配置、成本和结果的历史数据。", "side_effects": "引入控制面复杂度，冷启动与长尾任务可能选错团队。" }, { "claim_id": "R04", "type": "recommendation", "title": "让工件流动，减少聊天室广播", "text": "用带 schema、owner、版本和验收状态的工件传递；agent 订阅最小必要上下文，状态转移需校验。", "strength": "conditional", "scope": "长任务、并行任务与高上下文成本工作流", "supporting_evidence_ids": [ "E04", "E12", "E19" ], "contradicting_evidence_ids": [], "limitations": "过度压缩可能丢失弱信号。", "status": "validated", "show_in_summary": false, "evidence_basis": "结构化中间产物、里程碑和持久文件系统均提供正向证据。", "assumptions": "工件格式可演化且保留原始来源指针。", "side_effects": "需要 schema 迁移、冲突解决与额外存储治理。" }, { "claim_id": "R05", "type": "recommendation", "title": "Critic 必须独立且能看原始证据", "text": "执行者不能自报完成；Critic 应有独立 source access、反例检索、可执行测试和拒绝权。", "strength": "conditional", "scope": "高风险或可客观验收的工作", "supporting_evidence_ids": [ "E02", "E13", "E14", "E22", "E26" ], "contradicting_evidence_ids": [], "limitations": "多个同源 judge 仍可能共享偏差；独立 Critic 的净增益需按任务单独消融。", "status": "validated", "show_in_summary": false, "evidence_basis": "失败 taxonomy、专家稀释和信息整合失败支持独立复核；MAST 仅提供组合式 verifier/role/prompt 干预，不能证明独立 Critic 单项因果。", "assumptions": "Critic 与执行者有不同信息路径或工具。", "side_effects": "延迟和成本增加；需防止无限审查循环。" }, { "claim_id": "R06", "type": "recommendation", "title": "把共享记忆做成分层、可撤销的资产", "text": "区分 episode/team/workspace 记忆；记录 provenance、ACL、TTL、删除和 promotion；未经复核内容不进入长期知识层。", "strength": "contested", "scope": "跨 agent、跨 session 的 Multica workspace", "supporting_evidence_ids": [ "E19", "E21" ], "contradicting_evidence_ids": [], "limitations": "ACL 与 promotion 策略仍缺跨系统验证。", "status": "qualified", "show_in_summary": false, "evidence_basis": "持久文件工件有效，但权限化共享记忆仍是新兴方向。", "assumptions": "平台能追踪写入者、来源版本与消费记录。", "side_effects": "治理成本、遗忘/删除冲突和检索延迟增加。" }, { "claim_id": "R07", "type": "recommendation", "title": "仪表盘同时记录结果、轨迹与成本", "text": "最少记录成功/闭环、里程碑、token、延迟、调用、重试、人工接管和 MAST 类 failure labels。", "strength": "strong", "scope": "生产与试验环境", "supporting_evidence_ids": [ "E12", "E13" ], "contradicting_evidence_ids": [], "limitations": "failure label 需要人工抽检，不能当因果结论。", "status": "validated", "show_in_summary": false, "evidence_basis": "协调高不等于任务成功；失败分布随系统而异。", "assumptions": "trace 可采集且隐私/安全允许。", "side_effects": "存储与标注成本上升，指标可能诱发局部优化。" }, { "claim_id": "R08", "type": "recommendation", "title": "用敌对场景定位最早决定性错误", "text": "评测覆盖错误同伴、信息缺失/冲突、长上下文、重复步骤、错误终止、恶意动作；定位首个不可逆错误而非只给总分。", "strength": "strong", "scope": "发布前与持续回归 QA", "supporting_evidence_ids": [ "E13", "E20", "E22", "E23" ], "contradicting_evidence_ids": [], "limitations": "攻击与故障集合会随工具和模型演化。", "status": "validated", "show_in_summary": false, "evidence_basis": "错误同伴可说服正确 agent，14 类 failure modes 可分离观察。", "assumptions": "任务 trace 与状态转移可重放。", "side_effects": "用例维护成本高，可能暴露内部防护细节。" } ], "validation": { "critic_checks": [ { "claim_id": "C01", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E01", "E03", "E10" ], "no_contrary_note": null, "revision_reason": "检索 single-agent、best-of-n、matched-budget 与正向 MAS；保留‘取决于’而非‘普遍失败’。" }, { "claim_id": "C02", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E24" ], "no_contrary_note": null, "revision_reason": "检索 role/SOP ablation 与 single-agent workflow replay；同质 workflow 可折叠，因此限定为接口与工件价值。" }, { "claim_id": "C03", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E09", "E17" ], "no_contrary_note": null, "revision_reason": "检索 agent count、diversity、expert reveal、zero-cost cooperation、distributed information 与 topology scaling；保留可运行/部分正收益，否定单调规模律。" }, { "claim_id": "C04", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E20", "E22", "E23", "E26", "E27" ], "no_contrary_note": null, "revision_reason": "检索 debate、consensus、persuasion、expert dilution、distributed information 与 matched-budget；删除未经必要条件实验支持的‘只在’，改为观察性范围。" }, { "claim_id": "C05", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E17" ], "no_contrary_note": null, "revision_reason": "检索 equal token/call budgets、single scaling、best-of-n 与复杂 MAS；补入直接配平证据后仍因任务范围将 strong 降为 conditional。" }, { "claim_id": "C06", "outcome": "qualified", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E24" ], "no_contrary_note": null, "revision_reason": "检索 routing/search/early-stop 的离线成本与同质 workflow 折叠；要求重复任务可摊销并保留 simple baseline。" }, { "claim_id": "C07", "outcome": "qualified", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E16" ], "no_contrary_note": null, "revision_reason": "检索 structured artifact、free chat、latent communication 与 auditability；latent state 可能更高效，因此结论限定为可控性而非绝对性能。" }, { "claim_id": "C08", "outcome": "no_direct_contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 shared memory、ACL、provenance、deletion 与 memory utility；未发现否定治理必要性的直接反证，因证据新仍维持 contested。" }, { "claim_id": "C09", "outcome": "no_direct_contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 outcome-only、coordination score、trace/cost/failure evaluation；Silo-Bench/HiddenBench 补强过程诊断，未发现支持单指标充分性的直接反证。" }, { "claim_id": "C10", "outcome": "no_direct_contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [], "no_contrary_note": "在上述范围与截止日内未发现直接反证；这不等于不存在反证或已系统穷尽。", "revision_reason": "检索 hallucination-only、cooperation failure、communication-reasoning gap 与 system-level failure taxonomy；未发现将多 agent 故障完整归为底模幻觉的直接反证。" }, { "claim_id": "C11", "outcome": "contrary_found", "scope": "arXiv/ACL/ICML/ICLR/NeurIPS 官方来源；2023–2026 奠基集与 2025-01-15 至 2026-07-15 前沿窗口；按 claim 主题检索正反结果、简单基线和外推边界。", "finding_evidence_ids": [ "E23", "E24", "E28" ], "no_contrary_note": null, "revision_reason": "检索 latent/search/memory 新机制的 simple baseline、audit 和 matched-budget；直接 LatentMAS 复核否定 learned alignment 的稳定必要性，因此收窄前沿判断。" } ], "metadata_audit": { "status": "passed", "record_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/c7edc113ac506e78bb422e23d943be50b388d62c/research/agent-teams-2026/metadata-audit.v2.json", "source_count": 27, "required_fields": [ "source_kind", "title", "authors", "version", "submission_year", "venue_year", "publication_status", "venue", "track", "official_url", "venue_url" ] }, "consumer_attestation": { "status": "passed_release_v2_trial", "consumer_identity": "/root/manifest_consumer_v2_release5", "prior_involvement": "none", "manifest_version": "2.0.0", "snapshot_digest": "sha256:1cab26e51999310225fb08e05621ddfdbcad7ca3e478bc37181af0d614484a8c", "input_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/2aeb71cf4388a238e76a95ac6d6c715ab1c9dd3c/site/reports/agent-teams-2026/data/evidence-snapshot.v2.json", "attestation_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/4c97d4365ae73e41e9ce8db4d118fa242e48e41d/research/agent-teams-2026/consumer-attestation.v2.json", "attestation_digest": "sha256:92795e0f2a423958148e52d8d536c3290c2c7de4a0239db6c6b8549bb52fa4de", "transcript_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/4c97d4365ae73e41e9ce8db4d118fa242e48e41d/research/agent-teams-2026/manifest-consumer-validation.v2.md", "scope": "immutable evidence snapshot only; no report, research notes, Git history, issue, conversation history, paper originals or internet" } }, "outputs": { "report_url": "https://liaoyuanning.github.io/paper-learning-library/reports/agent-teams-2026/", "manifest_url": "https://liaoyuanning.github.io/paper-learning-library/reports/agent-teams-2026/data/evidence-manifest.v2.json", "model_disclosure": "AI 辅助生成：GPT-5 Codex（Multica PM-Paper runtime）；Scout、Reader、Critic 分工；PM-Paper 综合与验收", "copyright_policy": "仅保存元数据、必要短证据定位、结构化提取和综合判断；不存储论文全文", "generated_on": "2026-07-15", "time_precision": "day", "release_state": "release_v2", "metadata_audit_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/c7edc113ac506e78bb422e23d943be50b388d62c/research/agent-teams-2026/metadata-audit.v2.json" }, "promotion_candidates": [ { "claim_id": "C01", "eligibility": "promoted", "durable_knowledge_id": "AGE-185", "issue_uuid": "c807b74e-b65f-4424-b2b9-38ada71b0aad", "issue_pointer": "mention://issue/c807b74e-b65f-4424-b2b9-38ada71b0aad", "promotion_reviewed_on": "2026-07-15", "next_review_on": "2026-10-15", "revalidation_triggers": [ "manifest is superseded or retracted", "same-budget cross-task counterevidence appears", "claim is applied to a production continuous-action workflow" ] }, { "claim_id": "C05", "eligibility": "promoted", "durable_knowledge_id": "AGE-186", "issue_uuid": "5e7a4332-0f1d-4e0c-831d-e6ed7431395f", "issue_pointer": "mention://issue/5e7a4332-0f1d-4e0c-831d-e6ed7431395f", "promotion_reviewed_on": "2026-07-15", "next_review_on": "2026-10-15", "revalidation_triggers": [ "benchmark changes the budget definition or comparison baseline", "cross-task matched-budget counterevidence appears", "evaluation target is a persistent-action system" ] }, { "claim_id": "C09", "eligibility": "promoted", "durable_knowledge_id": "AGE-187", "issue_uuid": "f0a25456-ddd4-46c1-9ae7-94e6ed43fbd4", "issue_pointer": "mention://issue/f0a25456-ddd4-46c1-9ae7-94e6ed43fbd4", "promotion_reviewed_on": "2026-07-15", "next_review_on": "2026-10-15", "revalidation_triggers": [ "a more reliable workflow evaluation method appears", "trace or cost availability materially changes", "the claim is revised by a later manifest" ] }, { "claim_id": "C08", "eligibility": "not_eligible", "reason": "共享记忆治理仍以产品推断和新兴预印本为主。" }, { "claim_id": "C11", "eligibility": "not_eligible", "reason": "前沿机制尚未积累跨研究复验。" } ], "report": { "claim_links": [ "C01", "C02", "C03", "C04", "C05", "C06", "C07", "C08", "C09", "C10", "C11", "R01", "R02", "R03", "R04", "R05", "R06", "R07", "R08" ], "counterintuitive": [ "即使明确告诉团队谁是专家，共识也可能稀释专家判断；受控实验的最大差距达到 41.1%。", "相同 thinking-token 预算下，强单智能体在多跳推理中常匹配或超过 MAS。", "同一模型组成的 workflow 可能由一个 Agent 顺序执行并复用 KV cache，而不损失效果。", "能力更强不等于更会合作：零成本协作中 o3 只达到最优集体表现的 17%。", "消息送达不等于联合推理成功；Silo-Bench 和 HiddenBench 均揭示分布式信息整合断层。" ], "timeline": [ { "period": "2023", "title": "角色与会话", "text": "CAMEL、Debate、ChatDev、MetaGPT、AutoGen 证明角色化会话和流程可被编排。" }, { "period": "2024", "title": "结构与动态网络", "text": "ReConcile、DyLAN、More Agents 与 MacNet 开始分离多样性、选人、简单采样和拓扑效应。" }, { "period": "2025", "title": "路由、评测与失败", "text": "MaAS、MasRouter、MultiAgentBench、MAST 把成本、过程评测和故障定位带入主线。" }, { "period": "2026", "title": "可治理的协作介质", "text": "LatentMAS、Scaling、Experts Back、matched-budget、OneFlow、More Capable、Silo-Bench 与 HiddenBench 同时推进效率机制和直接反证边界。" } ], "taxonomy": [ { "title": "角色", "text": "专业化、边界、权限与终止权。", "question": "角色是否拥有可验收的契约，而不只是 persona？" }, { "title": "通信", "text": "自由文本、结构化消息、拓扑与 latent state。", "question": "传递的是最小必要证据，还是全部聊天历史？" }, { "title": "协调", "text": "领导者、共识、里程碑与冲突解决。", "question": "谁能推进、拒绝、升级和终止？" }, { "title": "记忆", "text": "episode/team/workspace 状态与治理。", "question": "谁写入、谁读取、何时删除、如何追溯？" }, { "title": "计划", "text": "分解、并行、依赖、重规划与闭环。", "question": "中间状态能否检查并从局部失败恢复？" }, { "title": "批判", "text": "独立证据、辩论、验证与拒绝权。", "question": "Critic 与执行者是否共享同一偏差和上下文？" }, { "title": "动态组队", "text": "按查询选择角色、模型、工具、拓扑和预算。", "question": "router 是否有成本、漂移与失败回退遥测？" }, { "title": "规模与评测", "text": "数量、多样性、成本、任务闭环和 failure trace。", "question": "复杂团队是否超过等预算 single / best-of-n？" } ], "consensus_matrix": [ { "topic": "角色 / SOP", "consensus": "清晰接口、结构化工件与终止控制有条件改善阶段化任务。", "unknown": "开放式知识工作、真实团队和长期演化中的外部效度。", "claim_ids": [ "C02", "C07" ] }, { "topic": "辩论 / 共识", "consensus": "候选差异与交叉检查可增益；共识不是 correctness。", "unknown": "如何抵御从众、恶意 peer 与同源 judge 偏差。", "claim_ids": [ "C04" ] }, { "topic": "动态路由", "consensus": "选人和早停可改善任务阶段质量成本比。", "unknown": "搜索成本摊销、线上漂移、长尾冷启动。", "claim_ids": [ "C06" ] }, { "topic": "共享记忆", "consensus": "持久工件可突破 context 与 session 边界。", "unknown": "权限、污染、冲突、删除和 promotion 的统一基准。", "claim_ids": [ "C08", "C11" ] }, { "topic": "规模", "consensus": "agent 数量不是单调收益；任务结构和信息多样性更关键。", "unknown": "如何在线估计有效信息通道并控制全链路成本。", "claim_ids": [ "C01", "C03", "C05" ] }, { "topic": "评测 / 失败", "consensus": "必须联看结果、轨迹、成本和故障类型。", "unknown": "开放任务 judge 偏差、因果归因与跨系统可比性。", "claim_ids": [ "C09", "C10" ] } ], "method": [ "边界先于检索：前沿窗口固定为 2025-01-15 至 2026-07-15；v2 又窄召回合作失败、分布式信息整合和 latent alignment 反证。", "先宽后窄：machine-readable 候选池 49 篇；18 篇核心由 8 篇奠基/典型与 10 篇前沿组成。Diversity 与 FS-Researcher 转为扩展，ICML More Capable 和 ACL Silo-Bench 进入核心；HiddenBench 保留扩展并进入证据链。", "逐 claim Critic：C01–C11 均记录 counter_search 范围、outcome、发现/未发现和修正理由；未发现反证不被表述为已证明。", "一手元数据审计：S01–S27 的 title、authors、version、submission/venue year、publication status、venue/track 与 URL 逐项对照 arXiv、OpenReview 和官方 venue；测试要求 audit 与 manifest 完全一致。", "机器证据优先：网页从 manifest v2 渲染；claim → evidence → source 使用闭合 ID。digest-covered evidence snapshot 与独立 consumer attestation 分文件、分指纹，并固定到 commit SHA。" ], "episode_loop": [ "现有知识差距：能力与合作脱钩、分布式信息整合断层和 LatentMAS 直接反证未闭环。", "新检索：15 条主题 query + 49 篇 machine-readable 候选 ledger。", "综合：18 篇核心重排为 8 篇奠基/典型与 10 篇前沿；两篇 2026 顶会工作进入核心，HiddenBench 进入扩展证据链。", "质量门：全量元数据审计、逐 claim Critic、固定 locator、snapshot digest 与独立 attestation 三元组测试。", "资产化：报告、manifest v2、evidence snapshot、候选决策、query、审计和独立 consumer attestation/transcript。" ], "limitations": [ "知识截止为 2026-07-15；之后的 arXiv 修订、录用状态和代码发布未纳入。", "oo-arxiv 依赖的 oo CLI 不可用；arxiv-cli 对部分已知 ID 返回空结果，使用官方一手页面复核，已用 arXiv 与会议官方页逐篇补核，但不能声称系统综述级穷尽。", "现有 benchmark 偏数学、代码、QA 和软件工程；真实组织的激励冲突、权限、安全和长期学习证据不足。", "多篇早期研究使用 2023 年 GPT-3.5/GPT-4；其绝对数字不能直接外推到 2026 模型。", "开源代码不等于完全可复现；商业模型漂移、系统 prompt、随机性、并发和 token 预算都会改变结果。", "本文的 Multica 建议是论文证据与产品约束的推断，不是论文作者直接结论；每条均标注前提和副作用。" ] }, "selection_protocol": { "candidate_count": 49, "core_count": 18, "source_count": 27, "decision_enum": [ "included", "extended", "excluded" ], "note": "extended/excluded 均为未进入 18 篇核心；每条保留版本、状态、URL 与理由。" }, "promotion": { "state": "partially_promoted", "promoted_claim_ids": [ "C01", "C05", "C09" ], "not_eligible_claim_ids": [ "C08", "C11" ], "promotion_reviewed_on": "2026-07-15", "next_review_on": "2026-10-15", "bidirectional_pointer_status": "complete" }, "evidence_snapshot_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/2aeb71cf4388a238e76a95ac6d6c715ab1c9dd3c/site/reports/agent-teams-2026/data/evidence-snapshot.v2.json", "evidence_snapshot": { "status": "immutable_release_v2", "manifest_version": "2.0.0", "snapshot_digest": "sha256:1cab26e51999310225fb08e05621ddfdbcad7ca3e478bc37181af0d614484a8c", "immutable_url": "https://raw.githubusercontent.com/LiaoyuanNing/paper-learning-library/2aeb71cf4388a238e76a95ac6d6c715ab1c9dd3c/site/reports/agent-teams-2026/data/evidence-snapshot.v2.json" } }