Multi-agent debate is a popular way to boost the reasoning of large language models, but it is not always worth the compute. A new paper on arXiv argues that the benefits of such debate are highly regime-dependent, and proposes a confidence-guided adaptive debate framework to make the process more efficient.
The core idea is to let the models' own confidence determine how much debate is needed, rather than always running a fixed exchange. The authors note that interaction can amplify informative signals under corrective conditions, though the abstract cuts off before detailing the full theoretical or empirical results.
The paper is a reminder that more debate is not always better; knowing when to stop is as important as the exchange itself.