Multi agent debate improves reasoning with less cost
Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate
Artificial IntelligenceMachine Learning
Summary
Sometimes, having many AI agents debate can help find better answers, but only if done the right way. The authors explain that debate works well when it's used selectively and structured carefully. They introduce a new system called LASE that picks when and how agents should debate, using a leader who guides the discussion. This method achieves strong reasoning with much less computing effort than always debating or simply voting.
What this means in practice
- •For ai system builders: Create more efficient multi-agent AI reasoning systems by selectively engaging debates only when beneficial, reducing computational costs.
- •For automation engineers: Implement leader-adaptive communication protocols in decentralized systems to improve decision making with less resource use.
Authors
Seungdong Yoa, Ye Seul Sim, Suhee Yoon, Sanghyu Yoon, Dongmin Kim, Soonyoung Lee, Bumsoo Kim, Junhyun Lee
Abstract
Multi-agent debate has shown promise for improving the reasoning of large language models, yet recent theory suggests its benefits are highly regime-dependent. While interaction can amplify informative signals under corrective conditions, symmetric debate dynamics are neutral in expectation, often making majority voting preferable. We reconcile these views by arguing that debate is effective only when invoked at the right time and with appropriate structure. Based on this insight, we propose LASE: Leader-Adaptive Structured Engagement, a leader-centric multi-agent debate framework that selectively engages interaction only in non-neutral regimes. LASE introduces an asymmetric leader-supporter architecture that enables directed information flow and selective signal amplification, while defaulting to simple aggregation otherwise. Experiments across diverse reasoning benchmarks show that LASE achieves multi-agent-level performance with near single-agent token cost, substantially improving efficiency over static debate and voting baselines.