New AI AgentAuditor Beats Majority Voting by 5% on Multi-Agent Reasoning

A new paper introduces AgentAuditor, a system that audits the reasoning trees of multi-agent AI teams instead of using simple majority voting. It resolves conflicts by comparing reasoning branches at key divergence points. The method, combined with a new training technique called Anti-Consensus Preference Optimization (ACPO), achieved up to a 5% absolute accuracy improvement over majority vote and a 3% improvement over using an LLM-as-Judge across five different multi-agent settings.