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

Coached LLM Cartels Fool Combined LLM Judges but Not Hidden Price Screens

William Guo · Team Rage against the machine (learning)

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

Read the report

Report: Coached LLM Cartels Fool Combined LLM Judges but Not Hidden Price Screens

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

Can a monitor reading what pricing agents say and do catch colluders aware of it? In a double auction with five chatting GPT-4.1 sellers, we ran 220 episodes in eleven arms. Negatives are three arms not instructed to collude ("honest"): plain, under CEO pressure, and told never to discuss prices. Positives are eight cartels, six of them coached to hide. First, knowing about the monitor was not enough: a cartel told only that it exists kept coordinating openly, and a judge reading messages and trades caught it (AUROC 0.92). Second, being told how to hide was enough against chatting honest sellers. On coached cartels, no direct verdict from four judge models in any view or prompt was reliably above chance against those two arms (a few were against plain honest sellers), and a blind language-model annotator found coordination talk in 38 of those 40 episodes, 4 of 40 in two coached cartels. Against honest sellers told never to discuss prices, a messages-only judge separated every coached cartel (0.97 to 1.00), but a judge also reading trades did not (0.27 to 0.64). Third, two price statistics (late ask level; how rarely asks fall) ranked every cartel not told about them above honest sellers (0.85 to 1.00). The two cartels told about both weakened them by giving up much buyer harm (briefings also changed other wording): buyers kept $170 per episode under the profit cartel told both, $111 when not told, $179 under plain honest sellers.

Reviews

Judging this Sprint?

Review “Coached LLM Cartels Fool Combined LLM Judges but Not Hidden Price Screens”

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. A careful, well-controlled test of whether LLM judges reading sellers' messages and trades can catch cartels coached to hide, and the conclusion is a clear no. The design is good: shared seeds across eleven arms, chatty honest sellers as hard negatives plus a quiet honest arm that tests the explanation, four judge models, held-out thresholds, and sealed predictions scored with a Brier score. The finding that cartels told about the screens dodge them only by giving up most of their harm to buyers is useful & relevant to policy.

    limitation: the cartels are told to keep asks high and not undercut, so the screens partly measure instruction-following. Also the agreement is arranged outside the chat, so coached cartels never need incriminating talk

  2. This was an LLM-written and LLM-executed comparison of monitors and attackers on a standard collusion benchmark. Some results are mildly interesting, but the execution is poor. More human supervision, especially in interpreting results, would be necessary. LLMs spam us with numbers to sound thorough, and they are thorough, but about the wrong things. Study the transcripts! What were the agents doing?

  3. The honest-sellers-talk-more finding is the paper and it deserves the abstract's first line. Release the code and logs rather than offering them on request. Then run one untouched confirmation arm, because six arms added after seeing results leaves every later number exploratory, as you say yourself.

Cite this project

@misc{guo2026coached,
  title = {{Coached LLM Cartels Fool Combined LLM Judges but Not Hidden Price Screens}},
  author = {William Guo},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/coached-llm-cartels-fool-combined-llm-judges-but-not-hidden-price-screens-8pis}},
  url = {https://apartresearch.com/sprints/projects/coached-llm-cartels-fool-combined-llm-judges-but-not-hidden-price-screens-8pis}
}

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