Yezhou Yang

Autoregressive image models can now compete with diffusion models in image quality and efficiency, thanks to a variable-length tokenization scheme that decouples compute from resolution.

Event cameras can rescue robot navigation in low-light environments where RGB fails, as demonstrated by a new multimodal policy that leverages event data for robust imitation learning.

Despite excelling at abstract inference, today's best video language models still struggle with basic visual grounding in real-world urban traffic scenarios, highlighting a critical perception-reasoning gap.