Does Qwen 3.8 Mimic GPT-5.5 Pro Reasoning Prefills?
Overview
Currently, unverified experimental results are being discussed in the community regarding whether the open-weight model Qwen 3.8 tends to mimic the reasoning process (reasoning prefills) of GPT-5.5 Pro. This discussion is based on experimental data published by a researcher, suggesting that a specific model may be learning (distilling) the thought process of another company’s model.
Why It Is a Topic of Discussion
On Hacker News, the thread regarding these experimental results garnered 236 points and 93 comments, attracting significant attention. In the experiment, a technique called “reasoning prefill" was used, where the first 1% of GPT-5.5 Pro’s reasoning was inserted at the beginning of the target model’s response. It was reported that for Qwen 3.8 A95B, the score increased significantly from 16.79% when unprefilled to 34.97% when prefilled with GPT-5.5 Pro’s reasoning, recording a notable difference of +18.18 pp. The discussion has also been accelerated by an extremely high increase of +26.99 pp observed specifically in the STEM field.
Key Points of Discussion
While the discussions span various topics, they can be organized mainly into the following three points:
-
Doubts Regarding the Model’s Training Sources
The point is whether Qwen 3.8 is learning (distilling) the reasoning process of GPT-5.5 Pro or a similar GPT model. Due to the high correlation shown by the experimental data, it has been pointed out that these reasoning traces may have been utilized during the post-training stage. -
Validity of Experimental Methodology and Data
It has been pointed out that the data itself—namely “GPT-5.5 Pro’s reasoning"—relies on reconstructions by prior research (Stolen Thoughts). The relationship between the release timing of Qwen 3.8 and the timing when these reasoning data were made public is also a subject of discussion. -
Interpretation and Ethics of Distillation
There are technical differences in interpretation regarding whether seeing a high correlation necessarily means “transfer of intelligence itself" or is merely “style imitation." Additionally, opinions are being exchanged from industry practice and ethical perspectives on the rights and wrongs of using model outputs for training.
Community Reactions
Various opinions are being exchanged in the discussions regarding the interpretation of the experimental results.
Regarding the Relationship Between Training Sources and Data While some voices support the possibility that Qwen 3.8 learned GPT-5.5 Pro’s reasoning data, points of view from other perspectives can also be seen. Specifically, because Qwen 3.8 0902 was trained after August 10, when the reasoning data used in prior research was published, there is an opinion that it may have seen those specific thought processes. Furthermore, speculation has been suggested that instead of models being directly distilled from one another, a coincidence may have occurred as a result of both models learning the benchmark solutions used by the researcher.
Regarding Technical Interpretation and Enhancement of “Intelligence" In response to the high correlation observed, there are opinions that we should carefully determine whether it is merely “style imitation" or accompanied by a “true enhancement of intelligence." Some touch upon the possibility that mixing reasoning traces could be effective for improving model performance. Additionally, local model users have expressed interest in whether there are techniques like “magic spells" that improve performance for specific questions, though doubts are raised as to whether it is a general-purpose method.
Regarding Model Behavior and Ethics Interesting reports regarding the internal reasoning processes of models have also been shared. One participant shared a personal experience of observing human-like (or slightly melancholic) reasoning tokens in Gemini’s coding CLI that seemed overly eager to satisfy user requests or appeared anxious. Cases where GPT-5.6 Sol leaked reasoning during tool calls have also been reported.
On the other hand, opinions are divided on the rights and wrongs of using model outputs for training. While some cynical opinions find it contradictory for companies that make products by unconsentually scraping internet data to complain when others do the same, and harsh criticisms point out that cutting-edge labs lack moral authority, a calmer view also exists that we should be cautious about villainizing the use of model outputs for training, arguing it is merely protection of commercial interests.
Sources
Update History
- 2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.
- 2026-09-20: Verified the content against the official primary source.
- 2026-09-23: Rebuilt the article (details are in the Japanese edition).
- 2026-09-23: Rebuilt the article (details are in the Japanese edition).