ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning

ARMOR-MAD is a new training-free heterogeneous multi-agent debate framework designed to enhance large language model reasoning by treating debate as conditional computation. It incorporates Pre-debate Agreement Routing (PAR), Early Agreement Stopping Evaluator (EASE), and Semantic Outlier Detection (SOD) to optimize debate processes, achieving significant accuracy improvements on benchmarks such as MATH Level 5 (65.5%), GSM8K (96.5%), MMLU (90.0%), and MMLU-Pro (81.5%). This framework highlights the importance of model heterogeneity and agreement-based control in improving the efficiency and accuracy of multi-agent debate systems.