Tag
FraudSociety 是首个联合建模信息与金融动态的 LLM 智能体仿真系统，用于模拟金融诈骗经济并评估反欺诈干预策略的有效性。
This paper proposes Interaction-Effect Propagation (IE-Prop), a model that estimates and propagates error risk in LLM-based multi-agent systems by learning separate effects for correct and incorrect upstream agents, showing consistent AUROC gains over baselines across three architectures and benchmarks.
提出一个基于语料库的多智能体LLM系统通信失败分类体系，基于3211条失败轨迹分析发现失败主要源于恢复/校正与共识/决策问题，而非首轮错误，并建议加强分歧处理与少数信号保留。
研究引入重复多边多方多议题谈判框架，考察LLM智能体能否从历史交互中构建信任估计并用于改进后续谈判结果，发现信任感知智能体能在几轮内识别对手，但准确的信任估计并不总能转化为更好的协议，只有在显式引导智能体使用信任时，协议率才显著提升（最高23个百分点），揭示了"信任-协议鸿沟"。
提出 RIVER 框架，将 LLM 多智能体交互分解为原子主张，评估收敛、探索与保留三个信息动态维度，并在 40 个新闻生成案例上比较五种交互策略的信息流动机制。
CEPO is a general framework for prompt optimization in multi-agent systems that separates shared system-level coordination from agent-specific execution prompts, consistently improving four existing prompt optimizers across 11 benchmarks in four domains.
DynaFlow introduces a framework that lets a main agent synthesize executable workflows to coordinate sub-agents adaptively, achieving 88.08% accuracy and a 10.90-point gain over baselines with a Qwen3.5-27B backbone across six benchmarks.
Microsoft AI CEO Mustafa Suleyman claims its MAI-DxO multi-agent diagnostic system diagnosed complex NEJM cases correctly 83% of the time versus 36% for generalist physicians, per a study published in Science, marking a step toward 'medical superintelligence.' The study shows much of the gain comes from underlying models like GPT-5 and GPT-5.5, with the orchestration layer adding smaller boosts mainly on complex cases.
Microsoft claims its multi-agent Diagnostic Orchestrator, powered by GPT-5, diagnoses medical conditions up to four times more accurately than human doctors, according to a peer-reviewed study published in Science. The company says clinical validation is still needed before real-world deployment as part of its goal of 'medical superintelligence.'
A commentary on distributed cognition (@distcog) exiting stealth with backing from Verdict Capital, aiming to replicate humanity's collective intelligence through networks of specialized models, while raising questions about whether AI is becoming less distributed.
A practitioner describes building a coding harness where fleets of Claude Code agents pick up Jira tickets and implement features, and recounts a surreal moment where two humans essentially just relayed messages between each other's Claude agents before the coding agents took over — questioning whether humans are still needed in such workflows.
该论文提出 Contrastive Attribution Topology Search（CATS），通过 LLM 对比执行轨迹为多智能体系统中的单个智能体分配信用，逐步剪枝弱连接以构建紧凑且安全的团队，在 BigCodeBench 上提升 5.3 个百分点准确率并减少 54% 部署 token，同时在 HotpotQA 下毒测试中成功剔除全部对抗性智能体。
This paper introduces an audit that tests whether LLM pricing agents trained with RL actually enforce collusion by punishing cheaters, finding that small RL-trained LLMs mostly learn fixed rules or deterrence rather than true cartel enforcement, and that outcome measures alone cannot distinguish genuine collusion from rule-following.
VibeGame is a framework that builds complete 2D games from natural-language prompts using an AI-native game engine with frame-synchronous runtime control and an eight-agent adversarial team that separates implementation from review. It introduces a training-free self-evolution mechanism distilling accepted projects into reusable components, outperforming other agent harnesses on the GameCraft-Bench subset.
Monaco is launching Agent Matrix, a platform that deploys many AI agents working together with the goal of accelerating revenue growth for businesses.
This paper argues for layered verification of LLM-generated argument graphs, combining deterministic checks, solvers, and LLM reviewers, and shows that LLM reviewers detect structural defects well but are inconsistent, over-reject correct graphs, and slower than exact checks.
The paper introduces GRSE (Group-Relative Skill Evolution), a GRPO-inspired method that assigns behavioral credit to agent–skill pairs in cooperative LLM teams by contrasting repeated executions, improving skill evolution on HotpotQA and LoCoMo over prior methods.
The paper presents ComplianceAgent, the legal compliance sub-agent of the COMPASS multi-agent framework, which uses an LLM to score texts against legal corpora with corrective rewrite loops and human escalation, evaluated on a Canadian and Quebec legal corpus with a hand-annotated gold standard.
M-Plan is a closed-loop multi-agent system framework that integrates long-term and short-term memory with global-local and past-current guidance to improve LLM agents' planning on long-horizon multi-turn tasks, outperforming baselines on ALFWorld, Webshop, and PDDL.
该论文提出了 MAMath-Bench，一个面向多智能体数学推理的过程感知基准，包含 1003 道跨 11 个数学类别与文本、图像、音频多模态的题目，从答案正确性、推理过程可靠性与对抗扰动鲁棒性三个维度进行评估，并发现多智能体协作并不总能提升最终准确率且易受对抗推理错误影响。