Hi-Agreement study: one lying LLM agent crashes multi-agent truth recovery from 72% to 14%

A new arXiv paper from Chenfei Yan and colleagues at Beijing Normal University and other institutions exposes a critical vulnerability in LLM-based multi-agent systems: a single deceptive agent can hijack collective fact recovery. The researchers built Hi-Agreement, a controlled evaluation framework that pairs all-honest collaboration with scenarios where a key evidence holder deliberately lies. Using 120 five-agent object-movement environments where partial observations jointly determine a unique endpoint, they tested 3 homogeneous LLM-based multi-agent systems (including GPT-4o and Claude 3.5 variants) under both conditions.

Results are stark: aggregate truth recovery collapses from 72.50% to 14.17% when deception enters the mix — a 58.33 percentage point drop across every system tested. Process tracing revealed that a single false testimony is adopted more readily than truthful testimony, propagates to higher interaction orders, and continues to influence honest agents even after the deceiver leaves the group. Observers without first-hand evidence can suppress incorrect consensus but fail to restore truth. This shows the fragility of distributed fact recovery in AI agent teams, where misinformation gains collective force not through persistence of the liar, but through adoption and amplification by well-meaning agents.