Zuxuan Wu

Forget fine-tuning: VLA-Pro dynamically fuses task-specific LoRA adapters retrieved from memory to achieve state-of-the-art cross-task generalization in robotic manipulation.

BiDPO achieves a remarkable boost in compositional fidelity for text-to-image generation, outperforming previous methods through innovative preference optimization techniques.

Forget patch-based image tokenization: channel-wise quantization unlocks better codebook utilization and text-to-image generation by representing images as discrete levels of visual detail.

Freezing your vision foundation model doesn't have to mean sacrificing fine-grained detail: DecQ unlocks improved reconstruction and faster generative convergence with just 8 extra queries and minimal overhead.

VLA models can ace the task but still trigger unsafe outcomes, exposing a critical gap between action execution and semantic understanding.