Qwen 3.8 and GPT-5.5 Pro Reasoning Prefills Analysis

Qwen 3.8 shows strong alignment with GPT-5.5 Pro reasoning

Experimental results indicate that Qwen 3.8 A95B exhibits a significant increase in response overlap with GPT-5.5 Pro when its reasoning channel is prefilled with the first 1% of the teacher model's reasoning. This suggests that Qwen 3.8 may have been post-trained using reasoning traces from GPT-5.5 Pro or a closely related GPT model.

In a test across 45 problems (15 STEM, 15 non-STEM, and 15 synthetic puzzles), Qwen 3.8's overlap with the teacher's visible answer increased by 18.18 percentage points (from 16.79% to 34.97%) when provided with the reasoning prefill. This effect was most pronounced in STEM subjects, where the delta was +26.99 percentage points.

Performance Metrics by Category

Category	n	Unprefilled	GPT-5.5 Pro reasoning prefill	Delta
STEM	15	19.26%	46.24%	+26.99 pp
Non-STEM	15	20.62%	33.42%	+12.80 pp
Puzzle	15	10.49%	25.23%	+14.75 pp
All	45	16.79%	34.97%	+18.18 pp

Methodology: Reasoning Prefills as a Distillation Signal

The experiment utilizes a "reasoning prefill" technique to detect potential model distillation. The process involves generating two responses from a target model:

An ordinary, unprefilled response.
A response where the first 1% of a teacher model's (GPT-5.5 Pro) reasoning is inserted into the target model's reasoning channel.

Researchers then measure the overlap (mean of unigram, bigram, and trigram source recall) between the target model's first 100 tokens of the visible answer and the teacher's visible answer. A high positive delta—meaning the model's output becomes significantly more like the teacher's output when prompted with the teacher's internal reasoning—is used as a signal for distillation.

Comparative Model Analysis

While Qwen 3.8 showed a dramatic shift, other models tested in the same experiment showed minimal or negative responses to the GPT-5.5 Pro prefill:

Kimi K3: Had the highest baseline overlap (31.11%) but only a modest increase of +4.54 pp with the prefill.
Inkling: Showed a slight increase of +0.46 pp.
DeepSeek V4 Flash: Experienced a slight decrease of -1.17 pp.

The researcher notes that in previous experiments, Qwen showed little movement toward Claude Opus 4.8, contrasting with its strong movement toward GPT-5.5 Pro in this study.

Community Insights and Counterpoints

Technical discussion surrounding these results highlights several alternative explanations and nuances regarding the reasoning traces:

Potential Data Contamination

Some observers suggest the overlap may not be the result of intentional distillation but rather data contamination. One user noted that Qwen 3.8 0902 was trained after the release of a paper on "stolen thoughts" (August 10), and may have simply seen those specific reasoning traces during training.

Stylistic vs. Intelligence Transfer

There is a debate on whether these results indicate a genuine transfer of intelligence or merely a stylistic influence. As one commentator noted:

"While this result does imply there was some training on the reasoning trace and output of GPT 5.5, it doesn't tell us how much of the source of its training it was... And it doesn't tell us how much it is more a stylistic influence rather than being a genuine lifting over of intelligence."

Observational Evidence

Some users reported anecdotal evidence of similar reasoning patterns. One user observed that GPT-5.6 Sol's internal reasoning (leaked via tool calls) mirrored the phrasing and structure of Qwen 3.8 27B's reasoning blocks, specifically in geometry optimization problems, suggesting a shared lineage of reasoning traces.