Zhe Xu

We propose MindGYM, a thinking-centric data synthesis framework that injects cognitive traits into QA generation, enabling language and vision-language models to self-synthesize high-quality, low-variance data for efficient fine-tuning.