gist.github.com via Hacker News

Reasoning-Prefill Test Suggests Qwen 3.8 Was Trained on GPT-5.5 Pro Traces, Answer Overlap Jumps 18 Points

Alibaba OpenAI Open Source ai-business

Summary

An independent experiment posted September 9 shows that prefilling the first 1% of GPT-5.5 Pro's reasoning trace into Qwen 3.8 A95B raises its answer overlap with GPT-5.5 Pro from 16.79% to 34.97%, a +18.18 pp jump concentrated on STEM problems (+26.99 pp). The author frames the effect as evidence that Qwen may have been trained on GPT-5.5 Pro or a closely related model, contrasting with an earlier run where prefilling Claude Opus traces produced almost no shift. The gist trended on Hacker News with 117 points and 55 comments.