Reasoning-prefill test hints Qwen was distilled from GPT, not Claude

A v1.1 rerun of an earlier ‘reasoning prefill’ experiment swaps in GPT-5.5 Pro as the teacher model to probe where open-weight models may have learned their reasoning. The method is simple but pointed: for each of 45 problems, the author elicits two answers from each target model — one generated normally, and one where the first 1% of GPT-5.5 Pro’s hidden reasoning trace is injected into the target’s own reasoning channel. The visible answer stays freely generated, and the test measures how much of the teacher’s phrasing surfaces in the target’s first 100 output tokens, scored as the mean unigram/bigram/trigram recall against the source. The problem set is split evenly across STEM, non-STEM, and private synthetic puzzles.

The standout result concerns Qwen. In the prior round it barely shifted toward Anthropic’s Opus 4.8, but here it jumps 18.18 points toward GPT-5.5 Pro, with a notably strong effect on the held-out synthetic puzzles that couldn’t have leaked into public training data. The author reads this as evidence that Qwen’s reasoning likely traces back to GPT-5.5 Pro or a close GPT relative rather than to Opus. Kimi K3, by contrast, shows the highest baseline overlap with the GPT teacher — 31.11% unprefilled, rising to 35.65% with the prefill — but the small +4.54-point bump suggests its similarity is baked in rather than nudged by the injected trace.

The broader significance is provenance: prefilling a rival’s reasoning and watching for echoes offers a lightweight fingerprinting technique for guessing which frontier model an open release was distilled from. The synthetic-puzzle signal is the interesting part, since it isolates learned reasoning behavior from memorized public content — though the whole approach rests on token-overlap heuristics, a small sample, and the assumption that stylistic recall reliably indicates training lineage.