Reasoning Prefills on a Few Open Models, v1.1
Gist 1 month ago 7
The author reran a reasoning-prefill experiment by inserting the first 1% of GPT-5.5 Pro’s reasoning into several target models and then measuring how much of the teacher answer appeared in the target outputs’ first 100 tokens. The evaluation used 45 problems, and the largest source-recall gain was Qwen3.8 A95B with an increase of +18.18 percentage points. The results show Qwen moved much more toward GPT-5.5 Pro in this run, suggesting it may have learned from GPT-5.5 Pro (or a close variant) rather than from Opus, while Kimi K3 changed less (+4.54 points).