Qwen 3.8 Susceptible to GPT-5.5 Pro Reasoning Prefills

Evidence suggests the Qwen 3.8 A95B model may have been trained using data distilled from GPT-5.5 Pro. A recent experiment by researcher wsxiaoys indicates that the model's outputs align significantly more closely with the proprietary teacher model when prompted with its specific reasoning patterns.

Using a technique known as "reasoning prefills," the researcher inserted the first 1% of GPT-5.5 Pro's reasoning traces into the reasoning channel of several target models. For Qwen 3.8 A95B, this caused a significant jump in answer overlap with the teacher model, rising from 16.79% to 34.97%—an increase of 18.18 percentage points. The experiment, which spanned 45 problems across synthetic puzzles, non-STEM, and STEM categories, found the most dramatic effect in STEM problems. In those cases, the overlap surged by 26.99 percentage points, moving from 19.26% to 46.24%.

The Mechanics of Distillation Detection

This research builds upon the methodology detailed in the "Stolen Thoughts" paper, which describes an exploit used to recover encrypted Chain-of-Thought (CoT) reasoning from proprietary APIs. By extracting these hidden reasoning traces and using them as prefills, researchers can test if a model's post-training process involved distillation. Distillation occurs when a smaller or open-source model is trained on the outputs of a more powerful proprietary model to mimic its capabilities.

While other models were tested, the results for Qwen were distinct. For instance, Kimi K3 exhibited the highest overall overlap with GPT-5.5 Pro at 35.65% with prefill, but the prefill itself only contributed a modest 4.54 percentage-point increase. According to wsxiaoys, Qwen's lack of movement toward Opus 4.8 in previous tests further suggests it learned specifically from GPT-5.5 Pro or a closely related GPT model.

Industry Implications

These findings provide empirical evidence of the hidden dependencies that open-weights models often have on proprietary AI. While distillation is a common industry practice to accelerate the development of smaller models, the ability to detect it through reasoning-trace analysis exposes the origins of a model's intelligence. It suggests that the "reasoning" capabilities of some open models may be a reflection of the proprietary data they were fed rather than independently developed logic.

What Remains

As the industry moves toward more complex reasoning models, the effectiveness of prefill analysis will likely become a standard tool for auditing model provenance. It remains to be seen if the developers of Qwen will address these findings or if other leading open-weights models will show similar susceptibility to proprietary reasoning traces when subjected to the same testing.