Qwen 3.8 follows GPT-5.5 Pro reasoning prefills

	Qwen 3.8 follows GPT-5.5 Pro reasoning prefills(gist.github.com)
	232 points by wsxiaoys 22 hours ago \| 90 comments
	tl;dr: When prefilled with the first 1% of GPT-5.5 Pro's reasoning, Qwen3.8 A95B's output overlap with the teacher's answer jumped +18.18 points (including +27 on STEM), while DeepSeek V4 Flash, Inkling, and Kimi K3 barely budged. Since Qwen showed little response to Opus 4.8 prefills in a prior experiment, the author suggests Qwen may have been trained on outputs from GPT-5.5 Pro or a closely related model.
	HN Discussion: ↑Explains and contextualizes the methodology, noting the authors' prior work on CoT recovery ~Questions the methodology since the only available GPT-5.5 traces come from the stolen-thoughts leak that Qwen could have trained on ↓Proposes alternative explanation: both models may have trained on the same benchmark solutions rather than one distilling from the other ~Argues the result doesn't justify strong conclusions about distillation extent or nature, and defends training on model outputs ↓Directly challenges the logical leap from the data to the conclusion about GPT-lineage training