The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization

Multi-agent debate (MAD) systems improve LLM reasoning through iterative deliberation, but remain vulnerable to debate collapse, a failure type where final agent decisions are compromised on erroneous reasoning. Existing methods lack principled mechanisms to detect or prevent such failures. To address this gap, we first propose a hierarchical metric that quantifies behavioral uncertainty at three levels: intra-agent (individual reasoning uncertainty), inter-agent (interactive uncertainty), and system-level (output uncertainty). Empirical analysis across several benchmarks reveals that our proposed uncertainty quantification reliably indicates system failures, which demonstrates the validity of using them as diagnostic metrics to indicate the system failure. Subsequently, we propose a mitigation strategy by formulating an uncertainty-driven policy optimization to penalize self-contradiction, peer conflict, and low-confidence outputs in a dynamic debating environment. Experiments demonstrate that our proposed uncertainty-driven mitigation reliably calibrates the multi-agent system by consistently improving decision accuracy while reducing system disagreement.

When several AI assistants work together to solve a problem by debating, they sometimes reach the wrong answer, even when individual assistants seem confident. This breakdown, where group discussion leads the AI team astray, has been hard to detect or prevent. We noticed that struggling AI teams show telltale signs, much like a confused person who hesitates, contradicts themselves, or argues in circles. We measured three such signs: how often an AI changes its mind, how much the AIs disagree, and how shaky the final group decision is. These signals reliably flagged when the team was about to fail. Building on this, we developed a training method that teaches AI teams to stay steady and commit to confident answers. This matters because AI teams are increasingly trusted with high-stakes reasoning in areas like healthcare and education, where unreliable group decisions could cause real harm.