Wentao Hu

Humanoid video generation breaks out of short, reactive lip-sync loops into proactive, hour-scale 720p streaming that seamlessly handles live interruptions and dynamic reference shifts.

A learnable moderator can transform multi-agent debates into more efficient reasoning processes, outperforming traditional methods by reducing redundancy and enhancing evidence aggregation.

BEE reduces inference latency and tool redundancy, achieving state-of-the-art performance in visual perception while maintaining competitive reasoning capabilities.

Continuous latent visual representations can drastically improve multimodal reasoning efficiency and performance, outpacing traditional discrete output methods.

RCA boosts contextual faithfulness in LLMs by amplifying relevant signals without increasing computational costs or altering attention distributions.