Auditing Multi-Agent LLM Reasoning Trees Outperforms Majority Vote and LLM-as-Judge

2026-02-10

AI Summary

This work addresses the limitations of traditional majority voting in multi-agent large language model systems, which often overlooks the evidential structure of reasoning paths and is susceptible to shared hallucinations. To overcome these issues, the authors propose AgentAuditor, a framework that explicitly models inter-agent consensus and disagreement by constructing reasoning trees and replaces global voting with a local conflict verification mechanism. Furthermore, they introduce an Anti-Consensus Preference Optimization (ACPO) strategy that leverages minority correct reasoning paths to rectify collective errors. Evaluated across five mainstream multi-agent setups, the method achieves up to a 5% accuracy improvement over majority voting and outperforms LLM-as-Judge by 3%, effectively transcending the constraints of existing aggregation mechanisms.

Technology Category

• Multiagent Systems: Mechanism Design
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Knowledge Representation and Reasoning: Preferences

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Multi-agent systems (MAS) can substantially extend the reasoning capacity of large language models (LLMs), yet most frameworks still aggregate agent outputs with majority voting. This heuristic discards the evidential structure of reasoning traces and is brittle under the confabulation consensus, where agents share correlated biases and converge on the same incorrect rationale. We introduce AgentAuditor, which replaces voting with a path search over a Reasoning Tree that explicitly represents agreements and divergences among agent traces. AgentAuditor resolves conflicts by comparing reasoning branches at critical divergence points, turning global adjudication into efficient, localized verification. We further propose Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator on majority-failure cases and rewards evidence-based minority selections over popular errors. AgentAuditor is agnostic to MAS setting, and we find across 5 popular settings that it yields up to 5% absolute accuracy improvement over a majority vote, and up to 3% over using LLM-as-Judge.

Problem

• multi-agent systems
• reasoning trees
• majority vote
• confabulation consensus
• LLM reasoning

Innovation

• Reasoning Tree
• AgentAuditor
• Anti-Consensus Preference Optimization
• Multi-Agent LLM
• Evidence-Based Adjudication