DynaDebate: Breaking Homogeneity in Multi-Agent Debate with Dynamic Path Generation

Large Language Model-based Multi-Agent Debate (MAD) frameworks enhance reasoning and collaboration, but existing approaches suffer from agents adopting identical reasoning paths, leading to errors and degenerating into simple majority voting. DynaDebate introduces dynamic path generation and allocation, process-centric debate, and a trigger-based verification agent. This framework generates diverse solution paths, shifts focus to step-by-step logic critique, and objectively resolves deadlocks with external tools. Extensive experiments show DynaDebate's superior performance across various benchmarks, surpassing existing state-of-the-art MAD methods.