Several new research papers published on arxiv.org address a critical vulnerability in multi-agent large language model systems: the tendency for agents to converge on incorrect answers through “confabulation consensus” or “shared misconception.”
According to arxiv.org, researchers introduced AgentAuditor, which organizes agent outputs into a “Reasoning Tree” that explicitly represents agreements and divergences in reasoning, rather than relying on simple majority voting. The system resolves conflicts by comparing evidence at critical divergence points. Across four multi-agent frameworks and multiple reasoning benchmarks, AgentAuditor achieved improvements of up to 5% absolute accuracy over majority voting while remaining token-efficient.
A separate paper on arxiv.org describes R²-MAD (Remember and Reweight for Multi-Agent Debate), which addresses the problem where “when a majority of agents initially converge on an incorrect answer, the debate process tends to amplify rather than correct the error.” R²-MAD equips agents with experience memory from past debates and uses a debate-state-aware retrieval policy to calibrate concept priors. According to the researchers, R²-MAD “achieves consistent improvements over existing single-agent and MAD baselines” across various benchmarks.
Both papers were accepted to EMNLP 2026, with the AgentAuditor paper published in September 2026 and R²-MAD appearing in EMNLP 2026 Findings.