Qwen 3.8 follows GPT-5.5 Pro reasoning prefills

A study tested whether open-source models (Qwen 3.8, DeepSeek V4 Flash, Kimi K3, and Inkling) could be influenced by reasoning prefills from OpenAI's GPT-5.5 Pro on 45 problems across STEM, non-STEM, and puzzle categories. Qwen 3.8 showed the largest improvement, increasing from 33.92% to 54.50% source recall when prefilled with GPT-5.5 Pro's reasoning, suggesting it may have learned from or been trained on similar models.