Qwen 3.8 follows GPT-5.5 Pro reasoning prefills
summary
Providing a small fragment of a high-performing teacher model's reasoning process can significantly increase a target model's likelihood of reaching the same final answer. This effect varies widely across different models, suggesting that some models are more susceptible to 'reasoning prefills' than others. The results indicate that certain models may have been trained on data derived from specific high-end models. This method serves as a tool to analyze the relationship between different AI systems.
key points
- The experiment used GPT-5.5 Pro as a teacher model, inserting the first 1% of its reasoning into target models to see if they would mirror the teacher's final answer.
- Qwen3.8 A95B showed a massive improvement in accuracy when given a prefill, particularly in STEM and puzzle tasks, indicating a strong alignment with the teacher model.
- Kimi K3 demonstrated the highest overall overlap with the teacher's answers regardless of the prefill, but the prefill itself provided a smaller relative boost.
who should read it
AI researchers and developers interested in model distillation, training data provenance, and the mechanics of chain-of-thought reasoning.
discussion skeptical
consensus
Commenters are skeptical about the claim that Qwen3.8 was distilled from GPT-5.5, noting that the evidence relies on specific reasoning traces that became public after a research paper release. Some argue that the model may have simply trained on these public traces rather than through secret exfiltration. There is general agreement that the 'thinking' style of models can be manipulated or mimicked, making it a weak signal for proving distillation.
top insight
A user points out a critical flaw in the methodology: the specific 'stolen thoughts' used to test Qwen were published on the internet in August, and since Qwen 3.8 was trained after this date, the overlap could be a result of training on public web data. This suggests that the 'smoking gun' for distillation might actually be common training data contamination.
Generated by an LLM from the article and thread. Check the sources before relying on it.