Comment by aesthesia

Implanting a foreign CoT should drop the performance due to the reward-hacked CoT language mismatch, or in any case it will give replies different from the suspected teacher, which is precisely what happens here. However it's just a single datapoint, there were plenty of attempts to figure it out. Just about everything is different in those two model series, from writing patterns to CoT strategies. If you are familiar with modern guardrails and Claude's raw CoT (which is trivial to leak), you know it's entirely different from DeepSeek's, and any claim that they trained on the CoT is extraordinary and requires extraordinary evidence. They need to prove their claims, not vice versa.

For the contrast, you can see how actual CoT distillation looks like in practice in various versions of GLM: make a Google ToS-breaking request, and see how GLM 4.6 or 4.7 repeats Google's conditional prompt injections in full in their CoT (Gemini 2.5-3.0 only regurgitated small snippets, because they used something closer to a "chain of draft", but GLM reconstructed it from Gemini's CoT during distillation). GLM 5.3 repeats Anthropic's prompt injections and Claude constitution, word by word. That's how distillation looks like.