Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026London

Fine-tuning a monitor for AI control

Brandon Brozorio, Bart Jaworski, Jan Ebbing · Team finetuners

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Fine-tuning a monitor for AI control

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

We wanted to investigate the beneficial effect fine-tuning can have on a monitoring model. We evaluated Qwen3.5-9B as our trusted model and Qwen3.5-27B as our untrusted model. We generated Supervised Fine-Tuning (SFT) training data of the form (action_transcript, malicious_label) using our trusted model on some sample tasks on LinuxArena (as well as comparing to a baseline of public trajectories on the LinuxArena website), then fine tune a Low-Rank Adaptation (LoRA) adapter on our trusted model to reject more malicious actions. Unfortunately technical blocks prevented us from running evaluations on the fine-tuned model.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety 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 AI safety 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.

  1. The author(s) of this paper show a solid understanding of control and the key concerns with training a monitor. Love the idea of only training the monitor on honest/attack trajectories generated by T. It would be sick to see if this generalizes to attacks generated by U.

    The trusted models of the future will be at least as capable as Opus 4.6, and if you spend a bunch of test time compute with this T (many ideas, testing them, iterating) I imagine you could get attacks on par with what the U of the future would be able to generate.

    Cool idea, though ofc the authors weren't able to get any results.

  2. This paper provides a framework to compare a fine-tuned monitor to a base version (prompted). Due to technical issues, the results are incomplete.

    The approach is sound and documenting failures (unable to run evaluations on the Lambda Cloud) is valuable, especially in a hackathon setting where time is limited and those issues can be debilitating.

    I'd hope fine-tuning monitors to be(come) standard practice in practical settings. That's one of the things the LinuxArena project hopefully makes accessible to a wide range of organizations and researchers.

    The paper can serve as an entry point to the setting (AI Control, Monitoring, and Scalable Oversight) as it is very well written and understandable and provides a clear roadmap on how to run experiments (if the required hardware is available).

Cite this project

@misc{brozorio2026finetuning,
  title = {{Fine-tuning a monitor for AI control}},
  author = {Brandon Brozorio and Bart Jaworski and Jan Ebbing},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/finetuning-a-monitor-for-ai-control-nai0}},
  url = {https://apartresearch.com/sprints/projects/finetuning-a-monitor-for-ai-control-nai0}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026