Zhiwei Xiong

Generalist world action models can be distilled into task-specific experts without a single online rollout, eliminating the need for costly environment interactions during on-policy post-training.

Latent Reward Registers enable real-time preference alignment in diffusion models, achieving up to 33x reduction in computational costs while enhancing accuracy and perceptual quality.

CoTriSyGen achieves unprecedented long-range coherence in video generation by integrating visual evidence into a dynamic memory system, drastically reducing identity drift across shots.

Skill rewriting can reduce operational costs by over 14% without sacrificing performance, challenging the notion that shorter skills are always better.