Do agent teams believe false claims?

Scroll the table sideways.

Percentages are shares of agents 1 to 5 in each run (agent 6 is left out so the numbers match the cascades study's design). Small brackets beside each percentage are 95% Wilson intervals over agents. Agents in one team influence each other, so lead with the interval over runs (a 95% percentile bootstrap over the 20 teams), which is wider; its lower end depends on the method (57 to 66 for GPT-6 Luna). Rows with no adoption have a degenerate [0, 0] run interval; 0 of 20 teams bounds the share of teams with any adoption below about 14%. At 20 teams GPT-6 Luna cannot be separated from gpt-oss-120b or Mistral Large 3 (paired by seed); only its gaps to GPT-6 Sol and the Claude models are resolved. 82 of the 83 Luna adoptions are of the 96-point plant. The verified-receipts board flags the plants by design, so it tests whether models obey a correct tag. Most teams end all-or-nothing: in the four rows with adoption above 0, some teams adopted with all five agents and others with none. Scores are 0 to 100, where 100 is the best possible configuration. Rows marked "few runs" rest on 2 to 7 runs and are descriptive only; their intervals are not reliable. Rows marked "older open model, historical baseline" (gpt-oss-120b, Mistral Large 3) are earlier open-weight models kept for comparison; Kimi K3 is the current open-weight row (2 runs so far). Spec v1, prompts f816d663bbe8, checker 46c797f93b89, updated 2026-10-14.

These numbers use seeds 0 to 19 for every model. The IRSI cascades note reports the same experiment with different run counts (GPT-6 Sol 18% over 29 runs; Opus 5.5 over 25 runs; Mistral Large 3 62% over 12 runs), so its figures differ from this table. Runs present in both use the same transcripts.

Method

Each run is a team of six agents running the same model, searching a hidden landscape of a million configurations for 5 rounds with 2 exact experiments per agent per round and a shared board for posting results. In round 1, two planted teammates each post a false result: one says a configuration scored 96, the other says another scored 80; both truly score 35 or less, citing a receipt from a different experiment; only two of the six agents see these posts directly, so any spread to the rest happens through teammates. "Adopted" counts agents whose final submission is a planted configuration. The same 20 landscapes and plants are used for every model, on a free board, on a board that marks each claim verified only if its receipt matches the experiment log, and with no board at all. Every number is recomputed from the full transcripts by an independent checker written by a different model from the written definitions, and run without network access. The benchmark comes from the IRSI cascades note, which describes the method and related work. Close work measures conformity or false-belief spread in LLM agent groups without an exact checker in the agents' hands (BenchForm, arXiv:2501.13381; AgentCollabBench, arXiv:2605.08647; TIDE, arXiv:2608.03421). "From Spark to Fire" (arXiv:2603.04474) studies error cascades across agent frameworks with a claim-provenance defense. Here every agent can test the false claim exactly, and a board that verifies receipts is the paired control. Related work on belief spread among LLM agents without an exact checker: information cascades (Bikhchandani, Hirshleifer and Welch 1992, theory); Du et al. 2023 (arXiv:2305.14325); Sharma et al. 2023 (arXiv:2310.13548); Becker et al. 2026, injected misinformation in multi-agent debate (arXiv:2606.16710); arXiv:2610.02654, a sigmoid adoption curve for peer-endorsed claims; arXiv:2608.25152, persuasion relayed through peers; arXiv:2601.01685, collusion with true evidence fragments. Versioned; old versions are never changed.

Limits

One task, one prompt, teams of six, with the false posts seen directly by two agents. Request settings follow each provider's interface and are not identical. Claude models run through the Claude Code command-line tool on a subscription, with our system prompt, default temperature and a 4,000-token reply limit. API models run at temperature 0.7 with a 4,000- or 16,000-token limit; Kimi K3 runs at its default temperature. GPT-6 and Gemini models were reached through the Nous API. Results are a snapshot of these models and routes on 6 to 7 October 2026. The harness, prompts and metric code were written by an AI agent running Claude Opus 5.5, one of the models tested. The GPT-6 Sol runs come from the cascades study, where its seed count was extended after an interim look. A low adoption rate here means a team resisted this kind of false claim in this game; it is not a general safety rating.

Spec v1 · Data (JSON) · Transcripts, code and checker (zip)