👥 Multi-Agent¶

💬 ACL2025 · 8 paper notes

📌 Same area in other venues: 📷 CVPR2026 (2) · 🔬 ICLR2026 (47) · 💬 ACL2026 (40) · 🧪 ICML2026 (24) · 🤖 AAAI2026 (26) · 🧠 NeurIPS2025 (17)

🔥 Top topics: Agents ×8

Beyond Frameworks: Unpacking Collaboration Strategies in Multi-Agent Systems: This paper systematically decomposes multi-agent collaboration into four dimensions (governance mode, participation control, interaction pattern, context management). Through extensive experiments on two context-dependent tasks, it demonstrates that the combination of centralized governance + instructor-controlled participation + ordered interaction + instructor summarization is optimal, reducing token consumption by up to 93% while maintaining or even improving accuracy.
CoMet: Metaphor-Driven Covert Communication for Multi-Agent Language Games: This paper proposes the CoMet framework. By integrating a hypothesis-testing-based metaphor reasoner and a self-improving metaphor generator, CoMet enables LLM agents to utilize metaphors for covert communication and semantic evasion in multi-agent language games. It significantly enhances the strategic communication capabilities of agents in Undercover and Adversarial Taboo games (improving the win rate from 0.20 to 0.70).
CortexDebate: Debating Sparsely and Equally for Multi-Agent Debate: This paper proposes CortexDebate, a multi-agent debate method inspired by the mechanism of the human cerebral cortex. By constructing a sparse dynamic debate graph and an evaluation module based on the McKinsey Trust Formula (MDM), it simultaneously addresses two core challenges of existing Multi-Agent Debate (MAD) methods: "excessively long input context" and "unequal debate caused by overconfidence."
DocAgent: A Multi-Agent System for Automated Code Documentation Generation: Proposes DocAgent, an automated code documentation generation system based on topological dependency sorting. Through a collaborative Reader-Searcher-Writer-Verifier workflow, it incrementally constructs context, significantly outperforming FIM and Chat baselines across completeness, helpfulness, and truthfulness.
GETReason: Enhancing Image Context Extraction through Hierarchical Multi-Agent Reasoning: GETReason is proposed as a hierarchical multi-agent framework that decomposes the context extraction of public event images into three sub-tasks: geospatial, temporal, and event. These tasks are collaboratively completed by specialized agents, achieving more accurate image context reasoning than existing methods.
Multi-Agent Collaboration via Cross-Team Orchestration: This paper proposes Cross-Team Orchestration (Croto), a scalable multi-team collaboration framework that organizes multiple independent agent teams for cross-team interaction, utilizing Hierarchy Partitioning and Greedy Aggregation mechanisms to fuse diverse solutions from various teams into superior results.
Preventing Rogue Agents Improves Multi-Agent Collaboration: A framework is proposed to detect "rogue agents" by monitoring agent uncertainty in real-time and to intervene accordingly. This framework achieves performance improvements of up to 17.4%, 2.5%, and 20% on the self-built WhoDunitEnv multi-agent collaboration environment, code generation tasks, and resource sustainability tasks, respectively.
Voting or Consensus? Decision-Making in Multi-Agent Debate: This work systematically compares 7 decision protocols (voting vs. consensus) in multi-agent debate (MAD). It is found that consensus protocols improve performance by 2.8% on knowledge tasks, while voting protocols improve performance by 13.2% on reasoning tasks. Two new methods, AAD and CI, are proposed to enhance answer diversity, yielding performance gains of 3.3% and 7.4%, respectively.