REASONING PREFILLS ON A FEW OPEN MODELS, V1.1

Summary from the TLDR Dev newsletter

This experiment seeds open models' reasoning channels with the opening portion of a GPT-5.5 Pro trace, then measures overlap with the teacher's visible answer. Qwen3.8 A95B showed the largest increase, suggesting that it may have learned from GPT-5.5 Pro or a closely related model.