Core Finding: Significant Imitation Detected in Qwen 3.8
Recent researcher wsxiaoys released v1.1 of the “Reasoning Prefills” experiment on Hacker News AI, focusing on Qwen 3.8 and Kimi K3 to measure how much open models imitate GPT-5.5 Pro’s reasoning style. The study is research-only with no product launch date, pricing, or availability information involved.
The experimental design inserts the first 1% of GPT-5.5 Pro’s reasoning output into the target model’s reasoning channel, while the visible answer remains freely generated. The similarity is then measured on the first 100 tokens of the target model’s visible answer, using unigram-bigram-trigram source recall averaging.
Key Metrics and Counterintuitive Results
The evaluation spans 45 problems: 15 STEM, 15 non-STEM, and 15 synthetic puzzles. Key findings include:
- Qwen’s substantial shift: Previously showing minimal movement toward Opus 4.8 in earlier experiments, Qwen jumped +18.18 percentage points in visible answer similarity under GPT-5.5 Pro prefill, especially pronounced on private synthetic puzzles.
- Kimi K3 lead but lowest gain: Kimi K3 achieved the highest overlap with GPT-5.5 Pro both without (31.11%) and with (35.65%) prefill, yet the increment was only +4.54 points, the smallest among reported models.
This reveals a notable reversal: Kimi K3 starts highest but is least responsive to the prefill intervention, while Qwen starts lower but shows dramatically stronger alignment. The data suggest Qwen may have learned from GPT-5.5 Pro or a closely related model, rather than from Opus.
Model Comparison Table
| Model | No Prefill Overlap | Prefill Overlap | Delta Change |
|---|---|---|---|
| Qwen | Not reported | Not reported | +18.18 |
| Kimi K3 | 31.11 | 35.65 | +4.54 |
Note: All values sourced directly from the article abstract. Other models lack reported numbers.
Practical Recommendations
- If you care about reasoning transparency and originality in AI outputs: This experiment demonstrates that inserting just 1% of a teacher model’s reasoning can significantly shift subsequent content, implying potential coupling between “free generation” and internal stylistic preferences.
- If deploying open models commercially: Qwen’s strong learning signal here warrants checking for GPT-series outputs in training data, especially for high-fidelity scenarios (e.g., exam solutions, standardized workflows). Consider adding similarity detection as a quality assurance step.
Final Thoughts
Reasoning prefill offers a novel mechanism to probe stylistic imitation in black-box models. Current evidence suggests cross-model reasoning migration may occur more readily than anticipated, shifting future debates from data leakage toward transparency in inference-time caching and generation mechanisms.