PEAR: New adaptive routing boosts multi-agent AI debate accuracy
Dynamically swaps agent roles to eliminate positional biases and improve reasoning...
Multi-agent debate has emerged as a powerful technique to improve LLM reliability through iterative peer critiques, but existing approaches suffer from persistent positional biases, amplification of unreliable agents, and high sensitivity to role assignments. A new paper introduces PEAR (Permutation-Equivariant Adaptive Routing Multi-Agent Debate) — an inference-time protocol that dynamically reconfigures which agents take on which roles and how they communicate across consecutive debate rounds. By strategically switching agent-to-role assignments based on evolving agent states, PEAR ensures no single agent can permanently occupy a privileged network position, distributing influence more evenly.
The researchers theoretically characterize PEAR as an equivariant sparse router: it preserves accuracy even when agents are relabeled, while reducing routing complexity and improving generalization. Comprehensive evaluations across four reasoning benchmarks (likely including math, logic, and commonsense) and six diverse LLM backbones show PEAR consistently outperforms the strongest existing debate baselines in average accuracy. The code has been released to enable further experimentation. This work addresses a critical flaw in multi-agent reasoning systems and could make AI debate more robust for real-world applications.
- PEAR dynamically reassigns agent roles and sparse topologies each debate round to prevent persistent biases
- Theoretically proven as an equivariant sparse router that maintains accuracy under agent relabeling
- Tested on 4 reasoning benchmarks across 6 LLM backbones, outperforming existing debate baselines
Why It Matters
Delivers more reliable, fair multi-agent AI reasoning by eliminating fixed role biases.