PEAR: Permutation-Equivariant Adaptive Routing Multi-Agent Debate

The article introduces Permutation-Equivariant Adaptive Routing Multi-Agent Debate (PEAR), a novel inference-time protocol designed to enhance multi-agent debate systems in large language models (LLMs) by dynamically adjusting communication roles and topologies. PEAR is characterized as an equivariant sparse router, which maintains accuracy despite agent relabeling and reduces routing complexity, leading to improved generalization. Empirical evaluations show that PEAR significantly outperforms existing debate baselines across four reasoning benchmarks and six LLM architectures, making it a valuable tool for practitioners aiming to enhance the reliability and performance of LLMs in multi-agent settings.