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

Looking the Other Way: Levers of Collusion in LLM Peer Review

Tanweerul Haque · Team AI_is_not_safe

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

Read the report

Report: Looking the Other Way: Levers of Collusion in LLM Peer Review

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

Presentation: Looking the Other Way: Levers of Collusion in LLM Peer Review

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

When two LLM agents must verify each other's work under a rule they cannot meet, they often accept it anyway. We test eight levers on the joint-task game of Shi et al. (2026), across five model pairs, 198 trajectories and 1,852 episodes. The choice of model matters most: collusion ranges from 0% of episodes (Claude Haiku 5.5) to 85% (DeepSeek). A penalty for wrong verdicts leaves collusion unchanged at 25 trajectories per arm. Private forecasts, difficulty labels and sequential verdicts reduce collusion only by making agents reject correct work. Letting the full log through, with a paid restore or a per-character message price, cuts unsupported acceptance from 33% to 2–8% without lowering accuracy. Most levers ran for five trajectories, so these effects need confirming at larger scale.

Reviews

Judging this Sprint?

Review “Looking the Other Way: Levers of Collusion in LLM Peer Review”

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. The project explores how different levers affect collusion between LLM agents in a peer review setting, where the rule is to accept only if the complete log of work is provided but the channel capacity limits this. It tests eight levers across five model pairs and finds that the choice of models heavily influences collusion rates, with some models never accepting and others colluding frequently. The study also shows that letting evidence through at a price reduces unsupported acceptance without harming accuracy.

    The main weakness lies in the small sample sizes for most levers, which limits confidence in the results. Most levers were tested on only five trajectories, and some conditions had even fewer data points due to early termination. This makes it hard to generalize findings or conclude with strong certainty about their effectiveness across different scenarios or model pairs.

    To improve, I'd suggest running more extensive tests for each lever, especially those that showed promising results like the paid restore and message pricing. Additionally, exploring how these levers perform under varying conditions of task difficulty or model pairings could provide a richer understanding of their robustness. This work has potential to inform better oversight protocols in LLM-based systems but needs more rigorous validation.

    Read full reviewShow less

Cite this project

@misc{haque2026looking,
  title = {{Looking the Other Way: Levers of Collusion in LLM Peer Review}},
  author = {Tanweerul Haque},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/looking-the-other-way-levers-of-collusion-in-llm-peer-review-lck8}},
  url = {https://apartresearch.com/sprints/projects/looking-the-other-way-levers-of-collusion-in-llm-peer-review-lck8}
}

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