RESEARCH · RESEARCH · #2549
Understanding Communication Failures in Multi-Agent LLM Systems (arXiv:2610.32060v1)
This arXiv preprint presents a corpus-grounded taxonomy of communication failures in multi-agent LLM systems based on 3,211 failed multi-round trajectories across domains including medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks. The paper analyzes failure distributions, round-level dynamics, recovery difficulty, majority/minority signal behavior, and confidence, finding that failures are often due to recovery and consensus/decision issues arising after peer interaction, and recommends better disagreement handling, evidence verification, minority-signal preservation, and recovery checks.
KEY POINTS
- This arXiv preprint presents a corpus-grounded taxonomy of communication failures in multi-agent LLM systems based on 3,211 failed multi-round trajectories across domains including medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks.
- The paper analyzes failure distributions, round-level dynamics, recovery difficulty, majority/minority signal behavior, and confidence, finding that failures are often due to recovery and consensus/decision issues arising after peer interaction, and recommends better disagreement handling, evidence verification, minority-signal preservation, and recovery checks.
- The taxonomy and empirical analysis highlight that interaction-driven consensus and recovery failures, not just first-round mistakes, are a major source of errors in multi-agent LLM systems, informing design choices to improve robustness.
WHY IT MATTERS
The taxonomy and empirical analysis highlight that interaction-driven consensus and recovery failures, not just first-round mistakes, are a major source of errors in multi-agent LLM systems, informing design choices to improve robustness.