Understanding Communication Failures in Multi-Agent LLM Systems

arXiv cs.AI Papers
multi-agent-systems llm-reasoning communication-failures taxonomy consensus failure-analysis

Summary

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。

arXiv:2610.32060v1 Announce Type: new Abstract: Multi-agent LLM systems aim to improve reasoning through debate, evidence sharing, and mutual correction, but interaction can also introduce failures that final accuracy alone cannot explain. Agents may converge too early, ignore dissent, suppress correct minority answers, or preserve a wrong framing across rounds. We present a corpus-grounded taxonomy of communication failures, built from 3,211 failed multi-round trajectories across medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks. Using this taxonomy, we analyze failure distributions across datasets and system designs, round-level error dynamics, recovery hardness, majority transitions, minority-signal suppression, and self-reported confidence. We find that failures are dominated by recovery/correction and consensus/decision problems rather than isolated first-round mistakes, with many emerging after agents first observe peer responses. Based on these findings, we recommend stronger disagreement handling, evidence verification, minority-signal preservation, and recovery checks before finalization.
Original Article

Similar Articles

FACTS-MAS: A Multi-Agent Architecture for Adaptive Forecasting from Heterogeneous Evidence

arXiv cs.AI

We introduce FACTS-MAS, a domain-agnostic multi-agent forecasting framework that decomposes heterogeneous evidence into specialized agent channels, validates and weights each via leakage-safe historical validation, and fuses them adaptively. Evaluated on Zillow housing inventory and FRED unemployment claims across 15 U.S. metros, it outperforms naive baselines and TimesFM, though a factor-adjusted exponential smoothing model matches its fused performance.

VerifiedLab: Harnessing Knowns for Unknowns

arXiv cs.AI

VerifiedLab introduces a multi-agent harness framework where AI scientists with distinct roles (PI, librarian, specialists, gate, critic) reason over field-verified records to turn scientific knowledge from passive artifacts into active systems, demonstrated on perovskite solar cells by recovering known trends and anticipating devices reported a year later.