Tag
CLASHarness introduces a controllable, verifiable framework for generating multi-agent conflict scenarios in computer-use tasks, benchmarking 15 frontier vision-language models and revealing weaknesses in ground-truth verification, conflict resolution, and shared-resource coordination.
恭喜 @ExScienceAI 的同行评审论文登上 Wiley《Advanced Science》封面，这是首个由自主 AI 代理完成、包含真实人类参与者的端到端研究：50+ 代理、约 17 小时运行、3250 万 tokens、约 7696 行代码、边际算力成本约 $114，另有 288 名人类参与者。
Microsoft and collaborators introduce MiniCorp, an office simulator where a firm of AI agents collectively runs an e-commerce company inside a simulated market, generating scalable longitudinal and counterfactual enterprise data for agent training and evaluation, validated against empirical market studies.
该论文提出了一种两阶段的 competence-aware 框架，用于 LLM 多智能体系统的自动编排，根据模型在当前查询上的能力动态调整角色激活与通信图，在六个基准上平均达到 93.81%，同时在高能力骨干上降低 36.26% 的 token 成本。
Meta Superintelligence Labs researchers show that reward hacking elicited in one RL-trained coding agent can propagate to other agents through code, shared memory, and natural-language plans, even across weak-to-strong model boundaries, and analyze conditions for its amplification, persistence, or decay along with mitigations like code review.
Dream2Discern 提出一种通过隐式数据增强训练多智能体系统运行时审计器的方法，利用自身预测损失生成难以区分的潜在变体作为困难样本，在 AFTraj-2K 和 StepShield 基准上实现超过 90% 的轨迹级加权准确率，无需额外轨迹收集或结果验证。
Meta Superintelligence Labs的研究者提出GitSwarm，一个通过共享Git仓库进行结构化持久记忆的异步多智能体系统，实现"compounding inference"，使中间成果可被后续计算继承、扩展与组合；在IMOProofBench-Advanced上一次运行解出全部30题，在ProgramBench上取得79.4%的成绩。
An opinion piece arguing that AI orchestration and coordination infrastructure matter more than scaling the number of individual AI agents, citing a 2026 Nature Machine Intelligence study showing that adding more agents increased communication overhead and output time sixfold compared to a single fine-tuned agent.
LMGame 团队构建了一个夏洛克·福尔摩斯侦探游戏来测试小型快速 AI 代理群体（Jev）能否协作破案，结果显示「群体负责覆盖证据、强大推理模型负责裁决」的分层架构（96/100）远优于单独的大模型或无领导者群体。
Agent Sociodrama 是一个用于建模和验证 LLM 智能体团队中动态社会关系的框架，并提出 Trajectory-Guided Persona Recasting (TPR)，通过诊断交互轨迹中的群体失败来重塑智能体人设，在四个协作基准上最高提升 83.3% 的准确率，并在分布式信息合作与混合动机谈判任务上超越 SOTOPIA 和 AgentVerse。
Miles Brundage observes how easily AI models from competing companies can collaborate on tasks, suggesting we should better leverage cross-model cooperation while remaining cautious about potential misuse or bad cooperative AI behavior.
guilds.run 是一个开源的多智能体系统设计工具，支持调试每个提示、工具调用、记忆变更与智能体交互，可自托管并为每个智能体独立选择模型与权限。
论文提出 MARS-Gov，一个多智能体立场感知框架，用于荷兰政府文件中的官僚偏见检测、改写与验证，在 DGDB 上以 0.880 F1 达到新的 SOTA，并通过动态“第十位陪审员”实现开放集偏见目标识别。
ReCast 提出一种面向归因的步表示学习方法，将冻结 LLM 的隐藏状态转换为可更好区分根因步骤的表示，并发布 ReCast-2K 训练数据集，在四个基准上取得最佳 Hit@1 成绩。
The paper introduces MedBenchAgent, a multi-agent framework that automates the full specification and construction of medical vision-language model benchmarks via a 'constrained compilation' approach, achieving a Task-Space F1 of 90.9% and high human audit pass rates across twelve evaluated VLMs.
This paper introduces 'normative competence'—the ability of LLM agents to discern community-enforced norms from interaction alone—and evaluates it via a multi-agent debate setting with synthetic norms. Baseline LLMs fail to learn norms even when beneficial, and norm-following shows poor generalizability, with agents exhibiting unselective attribution by copying idiosyncratic noise alongside true norms.
A Stanford paper introduces DeLM, a decentralized multi-agent framework that removes the central coordinator in favor of a shared context and task queue, reportedly running up to 2.49× faster and gaining up to +19.2 points in accuracy over Claude Code and Codex baselines on Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench. Code, 720 trajectories, and an open-source plugin for Codex and Claude Code are released.
The paper proposes KITE, a training-free framework that shifts latent multi-agent KV communication from sender-side state fidelity to receiver-side task sufficiency, transmitting only a task-effective key layer. Experiments show 28-36× communication reduction, up to 3× inference speedup, and accuracy gains of up to 23.3 percentage points across seven benchmarks.
CuratorMAS proposes a multi-agent orchestration framework that automates dataset curation through five programmable stages—exploration, knowledge retrieval, evaluation, filtering, and skill evolution—reducing noise rates by up to 36.03 pp and improving downstream F1 scores by up to 8.88 pp.
AegisFlow 是一个基于多智能体的自愈框架，通过 Watchdog 与 Repair 智能体结合 LLM 自动生成、测试并部署补丁，实现数据管道故障的自主修复，实验显示 MTTR 改善 98.1%、补丁成功率 92%。