New Multi-Agent System Framework Improves LLM Reasoning Accuracy Over Majority Voting

New Multi-Agent System Framework Improves LLM Reasoning Accuracy Over Majority Voting

Researchers have introduced AgentAuditor, a framework that improves how multi-agent systems aggregate large language model outputs by moving beyond simple majority voting. According to a paper published on arxiv.org, the system addresses a critical weakness called “confabulation consensus,” where agents with correlated biases converge on the same incorrect reasoning.

AgentAuditor organizes agent outputs into a Reasoning Tree that explicitly represents agreements and divergences in reasoning, then resolves conflicts by comparing branch-level evidence at critical divergence points. According to the research, this approach “consistently improves aggregation performance over majority voting, with gains of up to 5% absolute accuracy while remaining token-efficient” across four multi-agent frameworks and multiple reasoning benchmarks.

The researchers also introduced Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator to reduce conformity to misleading majority cues using evidence-verified preference supervision.

Separately, another arxiv.org paper addressing similar challenges proposed R²-MAD (Remember and Reweight for Multi-Agent Debate), which tackles the “shared misconception” problem where agents amplify rather than correct errors when a majority initially converges on an incorrect answer. This framework, accepted to EMNLP 2026 Findings, uses experience memory from past debates and dynamic retrieval mechanisms to calibrate agent reliability and modulate peer influence.