Staff, Alibaba Group
4 papers at NeurIPS 2025
Our empirically and theoretically informed method, which treats diversity as a reward, achieves new SOTA average performance across 7 benchmarks on SOTA LLMs with domain-undetermined data.
A scalable system for foundation model data processing, offering 150+ multimodal OPs, cloud-native efficiency (TB-scale on 10k+ cores), and diverse interfaces (Python/APIs/chat), widely adopted in research and industry (e.g., Alibaba Cloud).
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.
