Qwen 3.8 follows GPT-5.5 Pro reasoning prefills

AI Digest

Qwen3.8通过GPT-5.5 Pro推理预填充显著提升表现，尤其在STEM和合成谜题领域，可能学习了相关模型而非Opus 4.8。

Qwen3.8's reasoning performance surged by 18.18pp with GPT-5.5 Pro prefill, showing potential knowledge transfer from related models rather than Opus 4.8.

Key points

• 实验对比多个模型推理预填充前后的表现差异 Experiment compares model performance with/without reasoning prefill
• Qwen3.8在STEM和合成谜题领域提升最显著 Qwen3.8 shows largest gains in STEM and puzzle tasks
• Kimi K3与GPT-5.5 Pro重合度最高但增幅较小 Kimi K3 has highest overlap with GPT-5.5 Pro but minimal delta
• 数据暗示Qwen可能借鉴了GPT系模型而非Opus Data suggests knowledge transfer from GPT-like models
• 预填充技术有效提升模型推理输出质量 Prefill technique enhances reasoning output quality

Takeaway: 推理预填充技术能显著提升模型表现，揭示跨模型知识迁移的可能性。 / Reasoning prefill demonstrates effective knowledge transfer between models.

Why it matters 展示模型优化新方法，揭示跨模型知识迁移机制，对研发有实践指导意义。