MAIN CONTENT
How American answers become Chinese training material
The records reviewed by Reuters describe researchers prompting powerful AI systems for solutions, program code and step-by-step approaches, then selecting and refining their responses as training examples. This is often called teacher-student distillation: a larger or more capable model produces demonstrations, while a smaller model is trained to imitate useful patterns. The goal is not necessarily to copy a complete system but to preserve skills that can later be delivered without repeated calls to the original service.
A November 2024 Shanghai Jiao Tong University paper described selecting approximately 5,000 answers from OpenAI's o1, translating material using GPT-4o mini and assembling 10,750 bilingual examples to further train Qwen2.5-72B. Those figures describe the researchers' reported dataset, not evidence that the resulting system inherited all of o1's abilities. The authors themselves warned that an excessive reliance on distillation might discourage original advances.
Alibaba researchers reported a more structured pipeline in April 2025, generating multilingual material with GPT-4 and GPT-4o, then evaluating and filtering responses before training smaller Qwen models. A February 2026 paper involving Shanghai Jiao Tong University and Meituan described GPT-4o generating plans for teams of AI agents tackling programming tasks. Together these examples suggest a shift from isolated demonstrations toward repeatable methods for producing synthetic training data.
















