ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning
For LLM reasoning tasks, ARMOR-MAD offers a more accurate and efficient alternative to fixed-round multi-agent debate by dynamically controlling debate depth and agent contributions.
ARMOR-MAD introduces a training-free heterogeneous multi-agent debate framework that uses agreement-based routing and stopping to improve reasoning accuracy and efficiency, achieving 65.5%, 96.5%, 90.0%, and 81.5% on MATH Level 5, GSM8K, MMLU, and MMLU-Pro, respectively.
Multi-agent debate (MAD) can improve large language model reasoning, but fixed debate pipelines often waste computation and can amplify correlated errors among similar agents. We propose ARMOR-MAD, a training-free heterogeneous MAD framework that treats debate as conditional computation. ARMOR-MAD combines three components: Pre-debate Agreement Routing (PAR) decides whether independently generated Round-0 answers require debate; Early Agreement Stopping Evaluator (EASE) stops debate after convergence; and Semantic Outlier Detection (SOD) down-weights abnormal final answers during aggregation. Across MATH Level 5, GSM8K, MMLU, and MMLU-Pro, ARMOR-MAD consistently improves over fixed-round heterogeneous debate with the same model pool, reaching 65.5\%, 96.5\%, 90.0\%, and 81.5\% accuracy, respectively. The results suggest that genuine model heterogeneity and agreement-based control are both important for making MAD more accurate and efficient.