Alexander Panfilov(@kotekjedi_ml):As you might guess, this suggests that distilling reasoning traces may have been possible for a long time without ever breaking the cryptography. An anecdote: we find that prefilling Kimi-K3 reasoning with a few tokens of Opus reasoning measurably shifts its response toward Opus’s🤷‍♀️ A small memorization analysis showed that specific Claude and GPT reasoning spans are up to ~6 orders of magnitude easier to extract from Kimi-K3 than from the next-closest model.

Alexander Panfilov

@kotekjedi_ml

MATS 9.0 | PhD @ELLISInst_Tue & @MPI_IS doing AI Safety & Adversarial ML

Joined October 2013

376 Following 8.9K Followers

Alexander Panfilov@kotekjedi_ml

2026.08.11 12:00

As you might guess, this suggests that distilling reasoning traces may have been possible for a long time without ever breaking the cryptography. An anecdote: we find that prefilling Kimi-K3 reasoning with a few tokens of Opus reasoning measurably shifts its response toward Opus’s🤷‍♀️ A small memorization analysis showed that specific Claude and GPT reasoning spans are up to ~6 orders of magnitude easier to extract from Kimi-K3 than from the next-closest model.