Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate

Seungdong Yoa, Ye Seul Sim, Suhee Yoon, Sanghyu Yoon, Dongmin Kim, Soonyoung Lee, Bumsoo Kim, Junhyun Lee
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:149412-149429, 2026.

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-yoa26a, title = {Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate}, author = {Yoa, Seungdong and Sim, Ye Seul and Yoon, Suhee and Yoon, Sanghyu and Kim, Dongmin and Lee, Soonyoung and Kim, Bumsoo and Lee, Junhyun}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {149412--149429}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/yoa26a/yoa26a.pdf}, url = {https://proceedings.mlr.press/v306/yoa26a.html}, 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.} }
Endnote
%0 Conference Paper %T Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate %A Seungdong Yoa %A Ye Seul Sim %A Suhee Yoon %A Sanghyu Yoon %A Dongmin Kim %A Soonyoung Lee %A Bumsoo Kim %A Junhyun Lee %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-yoa26a %I PMLR %P 149412--149429 %U https://proceedings.mlr.press/v306/yoa26a.html %V 306 %X 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.
APA
Yoa, S., Sim, Y.S., Yoon, S., Yoon, S., Kim, D., Lee, S., Kim, B. & Lee, J.. (2026). Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:149412-149429 Available from https://proceedings.mlr.press/v306/yoa26a.html.

Related Material