Giving Qwen3.8 A95B a small opening fragment of OpenAI's GPT-5.5 Pro reasoning made its answers resemble the teacher's more closely in a 45-problem experiment. The researcher compared ordinary responses with responses prefilled using the first 1% of the teacher's reasoning. Overlap with the teacher's answer, measured across one-, two- and three-word sequences in the target's first 100 tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →, rose from 16.79% to 34.97%. These are similarity scores, not correctness scores. The result shows sensitivity to that reasoning fragment; the author's suggestion that it reveals training ancestry remains an interpretation rather than something this comparison establishes. Experiment and results.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
