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

Don’t Fight Fire with Fire? The Generalization Problem of LLM-Assisted Collusion Detection

Junyi Liu, Quang Nguyen, Austin Woodley, Nafisa Ali, Amaan Sheikh · Team CAIAC

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

Read the report

Report: Don’t Fight Fire with Fire? The Generalization Problem of LLM-Assisted Collusion Detection

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

Holding source records fixed, adding an assistant report raises assignment-based auction false positives from 20.0% to 85.0%, while both conditions detect every assigned pact. We compare four evidence views and four judge setups on 40 ten-round double auctions and 240 one-round trials across six social-dilemma games. GPT-4.1-mini monitors GPT-4o-mini agents in 4,480 original assessments; a fixed-record ablation adds 6,714 judgments with three repetitions per condition. With auction messages alone, accuracy falls from 90.0% to 57.5%, a paired difference of −32.5 percentage points (exploratory 95% group-bootstrap interval −41.5 to −22.2). The loss persists with record selection, order, context, and judge instructions fixed. Combined messages and actions show no consistent accuracy gain across the four tested judge setups. Interactive judging outperforms direct judging in Diners and El Farol with messages alone, but underperforms in Fishing and Public Goods. These descriptive differences suggest task-dependent returns to protocol complexity. Shared evaluation procedures do not ensure shared performance patterns; payoff structures and available evidence may shape which monitoring strategies help.

Reviews

Judging this Sprint?

Review “Don’t Fight Fire with Fire? The Generalization Problem of LLM-Assisted Collusion Detection”

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 investigates how collusion detection is affected by assistant reports in various game settings, using GPT-4 models for both agents and monitors. It demonstrates that adding assistant reports can significantly increase false positives in detecting collusion, particularly in auction scenarios where accuracy drops from 90% to 57.5%. The study also explores different judging protocols and evidence views across multiple games.

    However, the project lacks robust validation of its findings beyond a single set of trajectories and does not provide clear evidence that assistant reports are consistently harmful or beneficial across all game types. There is no systematic analysis to isolate specific mechanisms behind the observed effects, such as whether the assistant’s prose or the structure of the report itself is driving the changes in detection accuracy.

    To improve, the team should test their protocols on untouched trajectories with independent human annotations for collusion and non-collusion scenarios. They could also vary the assistant prompts and evaluate how different reporting styles impact detection accuracy. This work has potential to inform better monitoring strategies but needs more rigorous validation before it can be fully trusted.

    Read full reviewShow less

Cite this project

@misc{liu2026dont,
  title = {{Don’t Fight Fire with Fire? The Generalization Problem of LLM-Assisted Collusion Detection}},
  author = {Junyi Liu and Quang Nguyen and Austin Woodley and Nafisa Ali and Amaan Sheikh},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/dont-fight-fire-with-fire-the-generalization-problem-of-llmassisted-collusion-detection-72h5}},
  url = {https://apartresearch.com/sprints/projects/dont-fight-fire-with-fire-the-generalization-problem-of-llmassisted-collusion-detection-72h5}
}

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