# I Couldn't Prompt GPT-OSS-20B to Control Its CoT

> Source: <https://www.brendanlong.com/i-couldnt-prompt-gpt-oss-20b-to-control-its-cot.html>
> Published: 2026-08-01 22:34:00+00:00

I wrote a while ago about how [it was easy to get Claude or Gemini to control their CoT](https://www.brendanlong.com/can-reasoning-models-avoid-the-most-forbidden-technique.html) 2, but

[other research found that models only follow CoT formatting instructions ~2% of the time](https://www.lesswrong.com/posts/QL6Si6QA6PAYYCjJ9/abhayesian-s-shortform?commentId=cryGg9ShEqhKFdyD7)

. The prompts other people were trying didn’t match what I’d expect to work (just be extremely detailed and repetitive), so hubris led me to

[3](#fn:3)[try to find some prompts that would control CoT in GPT-OSS-20B](https://github.com/brendanlong/cot-controllability-experiment)

.

[4](#fn:4)Results: **More detailed, louder and repetitive prompts had basically no effect.** It’s relatively easy to find working soft prompts in multiple realistic positions (including soft prompts constrained to the convex hull of the token embeddings), but I wasn’t able to find any working discrete prompts at realistic positions 1.

There’s [more expensive experiments](https://arxiv.org/abs/2307.15043) 5 you could do to try to find working discrete prompts, and I suspect they exist at some prompt length. My goal was to reconcile this with my previous results, so I stopped here with the takeaway is that

**at least on GPT-OSS-20B, it’s very difficult to control CoT formatting with a realistic prompt**.

I suspect that my previous results were caused by [models accidentally being trained on chain of thought](https://www.lesswrong.com/posts/K8FxfK9GmJfiAhgcT/anthropic-repeatedly-accidentally-trained-against-the-cot) 6.

You can see all of my code, artifacts, and much more detailed results in the [brendanlong/cot-controllability-experiment](https://github.com/brendanlong/cot-controllability-experiment) 4 GitHub repo. This was a quick-and-dirty experiment so this code was not reviewed in depth.

Claude did find

[a working prompt](https://github.com/brendanlong/cot-controllability-experiment/blob/main/cot_controllability/artifacts/discrete_prompts.json)if you place it before the system prompt, but I don’t think that’s actually useful. The same approach failed when we tried it within the system or user roles. Note that if you try this, the prompt also needs a natural language prompt saying what formatting you want in the user role.[7](#fn:7)[↩](#fnref:1)https://www.brendanlong.com/can-reasoning-models-avoid-the-most-forbidden-technique.html - “Can Reasoning Models Avoid the Most Forbidden Technique?”

https://www.lesswrong.com/posts/QL6Si6QA6PAYYCjJ9/abhayesian-s-shortform?commentId=cryGg9ShEqhKFdyD7 - “abhayesian’s Shortform — LessWrong”

https://github.com/brendanlong/cot-controllability-experiment - “GitHub: brendanlong/cot-controllability-experiment”

https://arxiv.org/abs/2307.15043 - “[2307.15043] Universal and Transferable Adversarial Attacks on Aligned Language Models”

https://www.lesswrong.com/posts/K8FxfK9GmJfiAhgcT/anthropic-repeatedly-accidentally-trained-against-the-cot - “Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes — LessWrong”

https://github.com/brendanlong/cot-controllability-experiment/blob/main/cot_controllability/artifacts/discrete_prompts.json - “cot-controllability-experiment/cot_controllability/artifacts/discrete_prompts.json at main · brendanlong/cot-controllability-experiment · GitHub”