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

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

ARMOR-MAD is a training-free, adaptive framework for Multi-Agent Debate (MAD) in large language model reasoning, designed to overcome the computational waste and correlated errors of fixed debate pipelines. It integrates three core components: Pre-debate Agreement Routing (PAR) to determine if initial Round-0 answers require debate, Early Agreement Stopping Evaluator (EASE) to halt debate upon convergence, and Semantic Outlier Detection (SOD) for robust aggregation by down-weighting abnormal final answers. Across MATH Level 5, GSM8K, MMLU, and MMLU-Pro benchmarks, ARMOR-MAD consistently outperformed fixed-round heterogeneous debate, achieving 65.5%, 96.5%, 90.0%, and 81.5% accuracy, respectively, demonstrating the value of genuine model heterogeneity and agreement-based control.

Key takeaway

For AI Scientists and ML Engineers designing multi-agent LLM systems, relying on fixed-round, homogeneous debate is inefficient and risks amplifying correlated errors. You should prioritize genuine model heterogeneity and implement adaptive control mechanisms like pre-debate routing and early stopping. This approach, exemplified by ARMOR-MAD, significantly improves accuracy and computational efficiency, but ensure robust aggregation doesn't suppress correct "lone expert" minority opinions.

Key insights

Adaptive control and genuine model heterogeneity are crucial for efficient and accurate multi-agent LLM reasoning.

Principles

• Heterogeneity provides independent perspectives, reducing correlated errors.
• Adaptivity controls when debate is necessary, when it stops, and how answers aggregate.

Method

Heterogeneous agents (gpt-4o-mini, deepseek-v3, qwen-plus) generate Round-0 answers. PAR routes to debate if agreement < 0.67. EASE stops debate when all agents converge (phi=1.0) or after T_max rounds. SOD aggregates, down-weighting semantic outliers (lambda_out=0.7).

In practice

• Use diverse LLM families (e.g., gpt-4o-mini, deepseek-v3, qwen-plus) for agents.
• Implement agreement-based routing to skip debate for high-confidence initial answers.
• Employ semantic outlier detection to improve final answer aggregation.

Topics

• Multi-Agent Debate
• LLM Reasoning
• Adaptive Routing
• Model Heterogeneity
• Semantic Outlier Detection
• Conditional Computation
• Computational Efficiency

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning — ARMOR-MAD improves LLM reasoning efficiency and accuracy through adaptive, conditional multi-agent debate and outlier detection.
• Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation — Multi-agent LLM systems for ideation suffer diversity collapse due to interaction structures, not model insufficiency.
• MODF-SIR: A Multi-agent Omni-modal Distilled Framework for Social Intelligence Reasoning — A multi-agent, omni-modal framework uses distillation and TTA with LoRA for superior social intelligence reasoning.
• Critic-Guided Heterogeneous Multi-Agent Reasoning for Reliable Mathematical Problem Solving — A critic-guided heterogeneous multi-agent framework significantly improves LLM mathematical reasoning by enabling adaptive error correction and reducing reliance on large models.
• The impact of multi-agent debate protocols on debate quality: a controlled case study — Debate protocol design significantly impacts multi-agent LLM system behavior, balancing interaction richness and convergence.
• MADRAG: Multi-Agent Debate with Retrieval-Augmented Generation for Training-Free Analytic Essay Scoring — Multi-agent debate with retrieval augmentation offers a training-free, reliable approach to analytic essay scoring, outperforming prompt-based LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.