Reasoning Prefills on Open Models, v1.1

A short gist by wsxiaoys reruns an earlier experiment with a different teacher. The idea is to seed an open model's reasoning channel with the opening of another model's reasoning and see how far the open model's answer then drifts toward the other model's answer. A model trained on that teacher's outputs should recognize the trace and follow it; a model that never saw it should shrug it off. The earlier round used Opus 4.8 as the teacher and is linked from the gist but not in the vault; v1.1 uses GPT-5.5 Pro. It also cites "Stolen Thoughts" as a precursor.

Method

Each target model answers each problem twice: once normally, and once with the first 1% of GPT-5.5 Pro's reasoning inserted into the target's own reasoning channel. The visible answer is generated freely both times. The score is how much of GPT-5.5 Pro's visible answer shows up in the first 100 tokens of the target's answer, taken as the mean of unigram, bigram and trigram source recall. The set is 45 problems: 15 STEM, 15 non-STEM and 15 synthetic puzzles, the last described as private.

Results

Qwen's jump holds in every category: STEM goes from 19.26% to 46.24% (+26.99 pp), non-STEM from 20.62% to 33.42% (+12.80 pp), puzzles from 10.49% to 25.23% (+14.75 pp). The author's reading is that Qwen barely moved toward Opus 4.8 in the first round but moves strongly toward GPT-5.5 Pro here, including on puzzles that should not be in anyone's training data, so Qwen may have learned from GPT-5.5 Pro or a close GPT relative.

Kimi K3 is the other interesting row. It has the highest overlap with GPT-5.5 Pro whether prefilled or not, but the prefill adds little. High baseline overlap with a small delta and low baseline overlap with a large delta are different signals, and the experiment is built around the second one.

What it can and cannot carry

The design is reasonable: the delta compares a model with itself, so baseline differences in style and verbosity mostly cancel, and the same prefill goes to every model. The large, consistent Qwen effect beside three models that barely react is hard to explain by chance.

The limits are about scale and missing detail. Forty-five problems, one run per condition, no variance or repeated seeds reported. N-gram recall on the first 100 tokens measures surface wording, so it cannot tell "trained on this teacher's outputs" apart from "trained on data that shares its phrasing", such as a shared upstream dataset or another model distilled from the same teacher. The gist also does not say where the GPT-5.5 Pro reasoning text came from or whether it was a raw trace or a summary. The conclusion is phrased as "may have learned", and that is as far as the data goes.

Related

Two other stylometric instruments in the vault point at similar family resemblances by different routes. llm-cross-entropy-similarity compares how models write using character trigrams and opens on Kimi K3 paired with an Anthropic model. ai-comment-classifier found Kimi K2.7 and GLM so stylistically overlapping with every other model that their probability mass had to be redistributed, with its author guessing distillation as the reason. This prefill test differs from both: it measures how a model reacts to another model's reasoning rather than what its unprompted text looks like.

The allegation that open models are distilled from American frontier ones comes up in arguments-against-open-source-ai, which questions whether "Chinese model" even stays a stable category once distillation is involved. swe-1-7 and needle are the openly declared versions of teacher-student lineage, and kimi-k3 and deepseek cover two of the tested models. qwen3-8-flash-next covers the Qwen3.8 line around the third, including the Flash-Next preview of Qwen4's architecture, though not the A95B model tested here.

Larger evidence on the same question arrived later. The appendix of stealing-reasoning-traces-from-proprietary-llm-apis runs a full-scale version of this method and finds Kimi K3 and GLM-5.2 reacting unusually strongly to Anthropic reasoning, which its authors say does not prove distillation. That sits awkwardly with K3's small shift here, though the teacher (Opus there, GPT-5.5 Pro here) and the metric both differ. And anthropic-threat-report-september-2026 claims first-hand that Opus traces went into Qwen 3.5–3.7 as SFT data — earlier Qwen versions than the one tested here, and a claim the report gives no way to check. llm-distillation sorts these kinds of evidence by what each can and cannot show.