Qwen 3.8 follows GPT-5.5 Pro reasoning prefills
2 months ago
- The experiment tests the effect of prefilling the first 1% of GPT-5.5 Pro's reasoning into target models on their answer overlap with the teacher.
- Overlap is measured using unigram, bigram, and trigram source recall in the first 100 tokens of the answer.
- The evaluation includes 45 problems: 15 STEM, 15 non-STEM, and 15 synthetic puzzles.
- Qwen3.8 A95B showed the largest increase (+18.18 pp) in overlap with GPT-5.5 Pro, especially on synthetic puzzles.
- Kimi K3 had the highest baseline overlap with GPT-5.5 Pro (31.11%) but the smallest gain from prefilling (+4.54 pp).
- The data suggest Qwen may have learned from GPT-5.5 Pro or a closely related model, rather than from Opus.