Many LLM systems now rely on groups of agents that discuss, critique, and check one another, on the assumption that peers will catch individual errors. A new arXiv paper puts that assumption to the test against a minimal adversary—a single agent working against the group.

The study, described in the abstract, uses a ten-agent factual discussion task. The authors measure what they call 'hidden belief collapse': a situation where agents may privately hold wrong beliefs but do not express them in the discussion. This suggests that even if a group appears to reach consensus, individual agents might silently retain incorrect views.

Because the abstract is truncated, specific results and effect sizes are not available in the source. The paper's significance lies in its challenge to the uncritical use of multi-agent discussion as a reliability mechanism, and in offering a way to measure a failure mode that standard agreement metrics may miss.