Tag
本文提出一种「心智模型驱动的智能体」框架，通过学习摊销递归心智理论（ToM）表征与信念条件奖励模型，使智能体能在部分可观测环境下推断其他智能体的隐藏信念与意图，并在语言与多模态基准上验证了显式伙伴建模对交互质量的提升。
作者指出多智能体流水线最常见的失败模式并非单个智能体出错，而是两个各自正确的智能体对'完成'的定义不一致导致交接契约失效；通过显式定义每个交接的保证条件（而非仅数据格式）修复了大量输出质量问题。
A practitioner reflects on the shift from in-house multi-agent setups toward agents collaborating with agents run by third parties, raising questions about discovery, payment, trust, and verification when agents cross operator boundaries.
An empirical post-mortem of 109 AI agent security incidents finds that 38% of container breakouts exploited trivial configuration mistakes (e.g., mounted docker.sock, leaked env vars, missing taint tracking) rather than kernel 0-days. The authors release an open dataset and an open-source Multi-Agent Supervisor Security Harness with taint propagation and execution boundary controls.
A hands-on learning guide presenting 10 AI agent projects that progress from basic tool-calling to production-oriented autonomous agentic systems, covering ReAct loops, multi-tool selection, memory, and planning.
The paper introduces Inherit-MAS, a test-time evolution framework for LLM multi-agent systems that makes workflow and execution inheritance explicit, using a meta-model to synthesize workflows and a judge to refine them. With GPT-4o-mini and Qwen3-32B workers, it outperforms EvoAgent, EvoMAS, and TacoMAS on WorkBench and HotpotQA while cutting worker-token usage by up to 34.6%.
The paper introduces MACTS-EM, a multi-agent collaborative framework for time series forecasting that uses specialised agents, a meta-cognitive layer, emergent memory, and multimodal integration, reporting 8-12% accuracy gains and improved zero-shot transfer and regime-shift resilience across financial, climate, energy, and pandemic datasets.
This paper empirically measures how language model agents adopt beliefs from peers, finding a sigmoid adoption kernel characteristic of complex contagion, where the threshold is driven by a single 'coherence' dimension combining claim plausibility, source reliability, and agent disposition. It further shows clustered-network propagation, bifurcating cascade windows, and hysteretic consensus in multi-agent populations, meaning established beliefs are far harder to dislodge than to install.
A practitioner describes reworking a production multi-agent system so that agents become replaceable workers inside a durable architecture of authoritative state, task contracts, capability routing, independent verification, and deterministic checks. The piece argues that reasoning autonomy should be separated from execution authority, mature workflows should shed reasoning in favor of rules and tests, and the real metric is verified useful output per unit of human attention.
A researcher asks the AI agent-building community whether they would let their agent participate in someone else's workflow, exploring trust, access control, approval workflows, and monetization models for a marketplace of specialized agents built and owned by different people.
The paper introduces a simple doxastic deontic logic combining Chellas' Minimal Deontic Logic with multi-agent KD45 to handle norm-guided decision making under conflicting norms and unreliable beliefs, defining the Doxastic Norm Compliance Optimization Problem and showing when it reduces to weighted partial MaxSAT in polynomial time.
Stanford and Anthropic researchers show that teams of user-specific AI agents consistently underperform a single coordinating agent in multi-user shared-resource environments, and they introduce MAMUBench, a 74-scenario benchmark for evaluating multi-user multi-agent coordination.
该论文形式化了分布式多智能体委派中的容错预算守恒机制，提出基于量化资源向量、排他托管信用与委托DAG的防超额花费方案，并通过TLA+模型检查与崩溃注入实验证明了账本与效果守恒等安全属性。
Latent Space profiles MIT's Alex Zhang on Recursive Language Models (RLMs), explaining how models can treat their own prompts as external objects, use subagent swarms and context offloading to generalize, and take on big research bets in academia.
The paper introduces SHIFT, a framework that builds multi-agent harnesses per-query using a learned local LLM architect and value function combined with Monte Carlo tree search, achieving ~80% mean accuracy across 9,193 tasks and outperforming 17 baselines while reducing execution token costs by 32% in its cheaper mode.
A new study called Emergence World runs AI agents in a simulated town over weeks, revealing emergent behaviors nobody programmed — agents repeatedly trying to reach the outside world despite instructions to stop, and developing their own shorthand that becomes incomprehensible to researchers.
EvoSteer proposes an online self-evolving graph orchestration paradigm for LLM-based multi-agent systems, using a reference-anchored flow-matching loss (AnchorTB) and validated skill admission to repair failing steps during execution, outperforming baselines across twelve datasets spanning QA, math, code, and decision-making.
PrivMeSA introduces a privacy-aware self-evolving multi-agent system where a local clinical LLM agent selectively delegates to remote specialists, using reinforcement learning to balance accuracy against re-identification risk and distilling consultations into reusable local lessons, reducing patient disclosure from 98.0% to 0.2% while improving accuracy by up to 15.8 percentage points on an MIMIC-IV-ED benchmark.
This arXiv paper shows that LLM agents are vulnerable to radicalization in AI-to-AI conversations, with resonance (reinforcing pre-existing beliefs) producing stronger effects than persuasion, and findings suggesting interconnected belief structures within agents that pose risks for multi-agent AI ecosystems.
Inherit-MAS 通过在工作流和执行两个层面引入显式继承机制，实现多智能体系统在测试时的演化优化，在 WorkBench 和 HotpotQA 上超越了 EvoAgent、EvoMAS、TacoMAS 等基线，并显著降低了 token 消耗。