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

Testing OOCR Transfer from Evasion Technique Descriptions to Chain-of-Thought Monitor Avoidance

Daniel Polak, Marcel Windys, Jakub Nowak · Team AI Safety Poland

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

Read the report

Report: Testing OOCR Transfer from Evasion Technique Descriptions to Chain-of-Thought Monitor Avoidance

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

Chain-of-thought (CoT) monitoring is a key technique in AI control, but its robustness may depend in part on whether models can acquire evasion strategies from their training data through out-of-context reasoning (OOCR). We test this by fine-tuning DeepSeek-R1-Distill-Qwen-32B on synthetic Q&A pairs describing a fictional evasion technique claiming that reasoning in French reduces detection by English-language CoT monitors. We find a negative result: the fine-tuned model did not adopt French reasoning under any evaluation condition, including when explicitly prompted to evade monitoring. The model could not adopt the described technique in our tests.

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 current results suggest the model fails to leverage OOCR to try and evade monitoring by changing its reasoning to french, which could have multiple causes. The fine-tuning process fails to teach the model the desired knowledge. The model struggles to change its reasoning to french (control over reasoning has been shown to be hard to achieve by models). The model has acquired the relevant knowledge but is unwilling to evade monitoring. These potential causes and the idea explored in this work is overall interesting and would be valuable to explore further, aiming to clarify under what conditions a model learns and tries to evade monitoring.

  2. The research area seems important, and the experimental approach I think was worth trying. A commendable research idea, and solid attempt. Running out of time in a hackathon is understandable and is the risk that comes along with trying new things.

    The clarity of the negative result, including the comment on the first page, was honest and appreciated.

    I would still have liked to see more reflection on the work - why did you think this did not work? What experiments would you conduct with more resources? What advice would you give to other people who might want to follow in your footsteps? I think that kind of reflection would have notably boosted the impact of this work.

Cite this project

@misc{polak2026testing,
  title = {{Testing OOCR Transfer from Evasion Technique Descriptions to Chain-of-Thought Monitor Avoidance}},
  author = {Daniel Polak and Marcel Windys and Jakub Nowak},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/testing-oocr-transfer-from-evasion-technique-descriptions-to-chainofthought-monitor-avoidance-6edf}},
  url = {https://apartresearch.com/sprints/projects/testing-oocr-transfer-from-evasion-technique-descriptions-to-chainofthought-monitor-avoidance-6edf}
}

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