Lei Wang

Traditional frame-wise evaluation fails to capture the nuances of speech-driven facial motion, but a new sequence-alignment approach reveals clearer trade-offs in generative model performance.

Removing misleading reasoning from answer-correct CoT traces can drastically enhance fine-tuning performance in LLMs.

Quantizing rollouts in LLM RL pipelines introduces a training-inference gap that QaRL closes, leading to +5.5 performance on math problems.