Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning

DOI10.48550/arXiv.2606.29425MaRDI QIDQ7270398arXivFDO

Authors Yi Cai, Xiao-Yong Wei, Dayong Liang, Kaisong Gong, Changmeng Zheng

Publication date 28 June 2026

Published in arXiv

Tldr MoD is a unified framework for dynamic self-debate within a single model leveraging Mixture-of-Experts paradigm; it addresses three challenges: dual-routing, momentum switching, and unified self-debate to outperform both single-model baselines and conventional multi-agent systems in multimodal benchmarks with lower latency and reduced token consumption.

Full work available at URL https://doi.org/10.48550/arXiv.2606.29425

arXiv classification

cs.AIcs.CLcs.MAcs.MM

Cites work

• Attention Is All You Need

This page was built for publication: Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7270398)