Huadai Liu 刘华岱

I work on multimodal foundation models that perceive, reason and generate.

My research spans three parts of a multimodal foundation model: audio language models that generate speech, sound and music; multimodal LLMs that reason over video, audio and text; and post-training with reinforcement learning, on-policy distillation and chain-of-thought. Underneath them, I study the latent representations that diffusion and flow-based generators learn from, and samplers that make generation fast enough to serve.