Yang Luo

Distilling black-box video generators into autoregressive models doesn't require teacher scores or complex alignment—just cleverly paired rollouts and a discriminator.

Dialogue models can anticipate user intents and reduce redundant turns simply by injecting a lightweight intent-transition prior into the system prompt.

Achieve state-of-the-art 6D object pose estimation with 43x speedup by combining mask-aware pose proposals with amodal-driven refinement.

Forget balanced multimodal learning – letting the best modality lead the way actually unlocks better performance.

Solve the cold-start problem in CTR prediction by using multimodal LLMs to generate proxy embeddings that align with existing ID embeddings, enabling effective ranking of new items without usage data.

A confidence-based gating mechanism lets a 14B parameter reward model outperform 70B parameter models, achieving a new accuracy-efficiency Pareto frontier.