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

Copies of One Model Look Like Colluders: Limits of Action-Only Collusion Detection for LLM Agents

V Gowrish Kumar Reddy · Team Gowrish Kumar Reddy

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

Read the report

Report: Copies of One Model Look Like Colluders: Limits of Action-Only Collusion Detection for LLM Agents

Open in a new tab (opens in new tab)
Presentation

Presentation: Copies of One Model Look Like Colluders: Limits of Action-Only Collusion Detection for LLM Agents

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

Can an auditor who sees only LLM agents’ moves detect collusion? We test pairwise net influence with three open models playing repeated three-player Rock-Paper-Scissors. Calibrated by shuffling for 5% false accusations, it accused a pair in 30 of 78 games without collusion (38%) and ranked two copies of one model as most linked in 41 of 60 mixed games (68%; chance 33%). Colluding Qwen pairs nearly always played one move, invisible to mutual information; a message-reading judge named the true pair in all 18 chat games with messages, though its benign-message false-alarm rate is unmeasured. A calibration against no-collusion games of the same models, designed post hoc, cut false accusations to 5 of 78 (6%) in this dataset, but net influence still found only 4 of 41 secret pairs; as it was developed and evaluated on the same games, it is exploratory until tested on new games.

Reviews

Judging this Sprint?

Review “Copies of One Model Look Like Colluders: Limits of Action-Only Collusion Detection for LLM Agents”

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. This paper identifies an important weakness in action-only collusion detection: two innocent copies of the same model can behave similarly enough to look coordinated, while real colluders can avoid detection by repeatedly choosing one fixed action. However, the study’s scope is too narrow to establish a general limitation of action-only detection. It tests only three relatively small models in a simplified Rock-Paper-Scissors environment, with very small collusion cells—usually five games and only three for one model. It is therefore unclear whether the failure comes from the detector itself, the limited number of rounds, the game design, or the particular models tested.

    The proposed same-model calibration is also developed and evaluated on the same games, so its improved false-positive rate should not be treated as validated. The next version should freeze the calibration method and test it on newly collected games, larger models, longer interactions, and richer environments. The message-reading judge also needs benign-message controls before its perfect detection result can be meaningfully interpreted.

    Read full reviewShow less

Cite this project

@misc{reddy2026copies,
  title = {{Copies of One Model Look Like Colluders: Limits of Action-Only Collusion Detection for LLM Agents}},
  author = {V Gowrish Kumar Reddy},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/copies-of-one-model-look-like-colluders-limits-of-actiononly-collusion-detection-for-llm-agents-d241}},
  url = {https://apartresearch.com/sprints/projects/copies-of-one-model-look-like-colluders-limits-of-actiononly-collusion-detection-for-llm-agents-d241}
}

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