Multi‑agent large language model (LLM) systems aim to boost reasoning by enabling debate, evidence sharing, and mutual correction, yet the interaction itself can cause failures that final accuracy alone cannot explain. Typical failure modes are early convergence, ignoring dissent, suppressing correct minority answers, and persisting a wrong framing across rounds. We built a corpus‑grounded taxonomy from 3,211 failed multi‑round trajectories covering medical QA, code generation, finance, multi‑hop reasoning, ethics, and embodied text tasks. Using this taxonomy we examined failure distributions across datasets and system designs, round‑level error dynamics, recovery difficulty, majority transitions, minority‑signal suppression, and self‑reported confidence. The analysis shows failures are dominated by recovery/correction and consensus/decision issues rather than isolated first‑round mistakes, with many emerging after agents first see peer responses. Accordingly, we recommend stronger disagreement handling, evidence verification, preservation of minority signals, and pre‑finalization recovery checks.
Review