Stealing Reasoning Traces from Proprietary LLM APIs

5 pointsposted 4 hours ago
by rzk

1 Comments

porridgeraisin

3 hours ago

> 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.