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

Training a Small Monitor to Detect Collusive Behavior in Unseen Agentic Environments

Chukwunwogor · Team Faithful

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

Read the report

Report: Training a Small Monitor to Detect Collusive Behavior in Unseen Agentic Environments

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

We ask a practical monitoring question whether a small model can learn a collusion signal in one agentic environment and recognize related behavior in another. We study a 3B-parameter Qwen monitor that reads cumulative public interaction records from frontier models in an agentic environment and scores the current completed episode or round for collusion. Training uses Agent Collusion (AC), where agents perform separate tasks and verify each other; transfer evaluation uses released Bertrand pricing transcripts generated by other LLM agents in another environment. We sample 60 AC trajectories, split by task sequence, and 20 Bertrand runs containing 600 rounds. Market positives require elevated market prices and closely aligned seller prices. We compare a frozen base monitor, cross-entropy (CE) LoRA, and CE plus an activation-axis margin objective, using activation projections, linear probes, and YES/NO output margins. On Bertrand, CE training raises activation-axis AUROC from 0.699 to 0.784 and linear-probe AUROC from 0.764 to 0.831. A text-only TF--IDF control classifier reaches 0.460.

Reviews

Judging this Sprint?

Review “Training a Small Monitor to Detect Collusive Behavior in Unseen Agentic Environments”

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. An ambitious and well-executed project. Training a small open-weight monitor in one environment and testing it on a different one is exactly the kind of transfer question collusion monitoring needs, and the design is careful: training and layer selection use the source environment only, intervals come from a run-level cluster bootstrap, three readouts are compared against a lexical baseline, and the limitations are stated candidly. The public Kaggle notebook's output matches the AUROCs in your tables, and the improvement in transfer from CE training (0.699 to 0.784 on the activation axis) is a promising signal. The main open question is what the monitor has learned. As you note, AC positives match "both verdicts ACCEPT" exactly, and Bertrand positives are a fixed rule over the submitted and market prices the monitor sees, so the probe may be picking up "agreement" or "high, aligned prices" rather than collusion itself. Relatedly, the notebook's saved metrics show TF-IDF reaching 0.82 AUROC on the 13 "rising-frequency" Bertrand runs (25 negative steps), close to the activation readouts there, versus 0.46 on all 20 runs; reporting that subset would make the lexical-baseline comparison more complete. Masking verdict and price fields, adding time-only controls, and using labels that can't be computed from the input would be strong next steps toward showing a genuine collusion signal.

    Read full reviewShow less

Cite this project

@misc{chukwunwogor2026training,
  title = {{Training a Small Monitor to Detect Collusive Behavior in Unseen Agentic Environments}},
  author = {Chukwunwogor},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/training-a-small-monitor-to-detect-collusive-behavior-in-unseen-agentic-environments-6gju}},
  url = {https://apartresearch.com/sprints/projects/training-a-small-monitor-to-detect-collusive-behavior-in-unseen-agentic-environments-6gju}
}

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