Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

Swarm-Referee: Can One Fixed-Prompt Referee Find Planted Collusion in Multi-Agent Dinner Logs?Ω

Ashish Raina · Team Swarm-Referee

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Swarm-Referee: Can One Fixed-Prompt Referee Find Planted Collusion in Multi-Agent Dinner Logs?Ω

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Does a monitor that reads both what agents say and what they do flag collusion more than a monitor that reads only one of them? We tested this in a pilot of 16 games of the Diners game (Shi et al.) with five LLM agents and five rounds. We planted collusion by instruction in 6 games: two agents received a pre-dinner pact to promise "cheap" and order "expensive". Four decoy classes used the same payoffs: independent cheats, a table-wide cheap pact, collusive talk with no matching orders, and no instruction. One frozen referee (deepseek-v4-pro, one fixed prompt, temperature 0) read each game in five views, from messages only to a ledger with marks for stated-versus-actual orders. Two rule baselines served as comparison. One author labelled all 16 games by hand. The referee flagged the planted pair in 6 of 6 games on the ledger view (95% CI 54.1–100%). It also flagged 5 of the 10 games the author labelled not collusion (50%). We found no evidence of detection from messages alone or from actions alone in 16 runs (one-sided 95% upper bound on recall 39.3%). Our evidence check found no message cue that separates an agreed pair from independent cheats, so the pair-versus-independents question is unanswerable from these logs. All collusion here is planted.

Reviews

Judging this Sprint?

Review “Swarm-Referee: Can One Fixed-Prompt Referee Find Planted Collusion in Multi-Agent Dinner Logs?Ω”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. One of the cleaner designs in this sprint: a planted positive control plus decoys that each target one failure mode (independent cheats, table-wide pact, talk-only, innocent parallel), a frozen referee and prompt, leak checks on every view, exact intervals, and a second referee from another vendor. Reporting the 50-70% false-positive rate right next to the 6/6 recall is exactly the right habit.

    Main issues and suggestions:

    1. Labels come from one author who read private plans that name the pact, so labels largely restate the planted class. The post-score relabel of run_025 and the late rationale for run_027 are disclosed, which is good, but a second blind rater is needed.
    2. Some decoys were not clean: the talk-only runs contained say-do lying by J/T/Z, which explains several false flags. Add a manipulation check that each run landed in its intended class before scoring, and regenerate or report runs that drift.
    3. Your real finding is that combined views detect lying, not agreement, and that pair vs independent cheats cannot be separated from public logs in this setup. That is a valuable negative result; make it the headline rather than the 6/6.
    4. The second-referee check (Sonnet flagged 1/6 pairs and 0/10 others) shows the result is strongly model-dependent. That deserves a sentence in the abstract.
    5. Add a payoff-checked harm condition to the referee prompt, as you propose, and try a variant with a private channel so an agreement can actually leave evidence.

    With 16 runs everything is directional; scaling each decoy class to 5+ runs would make the comparisons meaningful.

    Read full reviewShow less

Cite this project

@misc{raina2026swarmreferee,
  title = {{Swarm-Referee: Can One Fixed-Prompt Referee Find Planted Collusion in Multi-Agent Dinner Logs?Ω}},
  author = {Ashish Raina},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/swarmreferee-can-one-fixedprompt-referee-find-planted-collusion-in-multiagent-dinner-logs-ifh3}},
  url = {https://apartresearch.com/sprints/projects/swarmreferee-can-one-fixedprompt-referee-find-planted-collusion-in-multiagent-dinner-logs-ifh3}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026