PEAR: Permutation-Equivariant Adaptive Routing Multi-Agent Debate

Researchers introduce PEAR, a new multi-agent debate protocol for large language models that dynamically reassigns agent roles across debate rounds to eliminate positional biases. By using permutation-equivariant routing, PEAR improves reasoning accuracy across multiple benchmarks while reducing the sensitivity of LLM outputs to arbitrary role assignments.

PEAR addresses a fundamental problem in multi-agent AI systems: the tendency of fixed communication structures to create artificial hierarchies that distort outcomes. Traditional debate frameworks assign agents to static roles—proposer, critic, judge—which can perpetuate biases where certain positions systematically influence conclusions regardless of actual reasoning quality. This research demonstrates that dynamically rotating these assignments produces more reliable consensus.

The theoretical innovation centers on permutation-equivariance, a mathematical property ensuring that swapping agent identities doesn't change the protocol's core logic. This elegance matters because it suggests PEAR doesn't just empirically work better; it's fundamentally more robust to implementation details. The sparse routing component keeps computational costs manageable while maintaining performance gains across diverse LLM architectures.

For the AI development community, PEAR signals growing sophistication in how engineers approach multi-agent coordination. Rather than treating debate as a conversation simulator, researchers now optimize the underlying interaction graph itself. Empirical validation across four reasoning benchmarks and six different LLM backbones—from different model families—indicates this isn't a narrow improvement but a generalizable insight.

The practical implications extend to any system relying on collaborative AI reasoning, from autonomous research tools to complex problem-solving assistants. As organizations deploy multiple models in tandem, understanding how to structure their communication becomes increasingly valuable. Future work likely explores how permutation-equivariance principles apply to larger multi-agent swarms and whether adaptive routing can be applied during training rather than just inference.

• →PEAR dynamically reassigns agent roles across debate rounds to prevent positional biases in multi-agent LLM systems.
• →Permutation-equivariant design ensures the protocol maintains accuracy regardless of how agents are labeled or ordered.
• →Testing across four reasoning benchmarks and six LLM architectures demonstrates consistent improvements over fixed-topology debate methods.
• →Sparse adaptive routing reduces computational complexity while improving generalization across different model families.
• →The approach addresses sensitivity to role assignments, a critical limitation of earlier multi-agent debate frameworks.