Assistant Professor, Shanghai Jiaotong University
3 papers at NeurIPS 2025
HoloV prunes 88.9% visual tokens in MLLMs while retaining 95.8% performance via spatial visual holistic context retention.
We propose a new paradigm to accelerate different components of the VLA model and achieved state-of-the-art performance.
We propose a progressive consistency distillation framework that enhances the efficiency of MLLMs by significantly reducing computational cost while preserving strong performance.
