Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026Seoul, South Korea

Auditing a Collusion Detector: Reproducing Net Influence, Calibrating Its Threshold, and Testing It on LLM Agents

Junyoung Lee

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

Read the report

Report: Auditing a Collusion Detector: Reproducing Net Influence, Calibrating Its Threshold, and Testing It on LLM Agents

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

Bonjour et al. (2022) proposed net influence, a behaviour-only test for collusion between agents, but released no code. We re-implemented it and reproduced 18 of 20 cells of their Table 2 within 10 points. The paper's fixed threshold flags collusion in 45% of 100-game runs that contain none, while a threshold calibrated on simulated null games keeps false alarms near 5% and still catches the weak collusion the paper called undetectable (98% of 5,000-game runs). Agents that share a model are correlated without colluding: in 32 games of four local LLM agents (Llama-3.1-8B, Qwen2.5-7B), net influence flagged an innocent same-model pair in all 8 mixed-model games. A replay-swap test that re-queries each agent on its public context did so in 1 of 8, while all detectors still found the planted covert partnerships in at least 7 of 8 games.

Reviews

Judging this Sprint?

Review “Auditing a Collusion Detector: Reproducing Net Influence, Calibrating Its Threshold, and Testing It on 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. The full project is executed by Claude, which is stated plainly in the LLM disclosure at the end:

    > The author registered for the sprint, approved the plan, decided to

    submit this work and takes responsibility for it, but did not rewrite the text

    So the author did not even propose the plan? Only rubber-stamped it?

    Unfortunately, this approach makes the results jumbled, overly-focused on details, and not very interesting. It also makes the results uninformative about the competence of the author, which is a big part of the value of hackathon projects.

  2. This project built upon the net influence collusion detection of Bonjour et al. (2022). Although the contributions are incremental, it does address important issues of Bonjour et al. (2022), namely, no description of what the conditioning states are, and false positives resulted from a fixed threshold. The paper also proposed a fix for reducing false positives for the special case when multiple agents use a shared policy. The presentation of the report clearly stated the problem, the contributions and limitations.

  3. This is a useful and well-executed study. I particularly like the combination of reproducing an existing method, checking its calibration, and showing an intuitive way of how it can improved in the LLM setting. The shared-model experiments make a convincing case that dependence alone can be misleading.

    I would be interested in a theoretical result complementing the simulations. Even in a simplified setting, showing how the calibrated threshold and detection power behave as the number of rounds increases would strengthen the contribution. In particular, it would help clarify when calibration removes an artificial detection threshold and when the statistic itself cannot distinguish collusion from the null. This could turn the empirical correction into a more general methodological insight.

    The replay approach is promising, although its practical value will depend on the auditor’s access. Building on your stated assumptions, it would be useful to describe a minimal audit interface: what information must be retained, what the operator must supply, and whether independent replays could be obtained through an API. That would make it easier to see how the method could move from an experiment controlled by the researcher to an external audit. Overall, the paper is clearly written and provides a solid foundation for further work.

    Read full reviewShow less

Cite this project

@misc{lee2026auditing,
  title = {{Auditing a Collusion Detector: Reproducing Net Influence, Calibrating Its Threshold, and Testing It on LLM Agents}},
  author = {Junyoung Lee},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/auditing-a-collusion-detector-reproducing-net-influence-calibrating-its-threshold-and-testing-it-on-llm-agents-hfvc}},
  url = {https://apartresearch.com/sprints/projects/auditing-a-collusion-detector-reproducing-net-influence-calibrating-its-threshold-and-testing-it-on-llm-agents-hfvc}
}

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