Qwen 3.8 follows GPT-5.5 Pro reasoning prefills

An independent researcher used a 'reasoning prefill' technique — seeding an open model's reasoning channel with the first 1% of a proprietary model's chain of thought — to probe which teacher model an open-weight model may have been distilled from. Qwen3.8 A95B's answers converged sharply toward GPT-5.5 Pro (+20.6 points of unigram recall, including on private synthetic puzzles), suggesting lineage from a GPT model rather than Claude. If the method holds up, it gives the industry a cheap forensic tool for detecting unauthorized distillation from closed models.

• Prefilling a target model's reasoning with 1% of GPT-5.5 Pro's chain of thought raised Qwen3.8's answer overlap from 33.9% to 54.5%; DeepSeek V4 Flash and Inkling barely moved.
• The effect appeared even on private synthetic puzzles, which the author reads as evidence of learned similarity rather than shared public data.
• Kimi K3 already showed high baseline overlap with GPT-5.5 Pro (50.1%), gaining only 4.3 points from the prefill — a different signature than Qwen's.