Minghao Li

Traditional aesthetic assessment models miss the mark by relying on static scores, while RED-Aes reveals the dynamic, comparative nature of human aesthetic perception for superior generalization.

LLMs can be sped up by over 2x without sacrificing accuracy, by compressing the input and predicting multiple output tokens at once using a unified framework.

Omni-modal RL post-training just got a whole lot faster: Relax delivers up to 2x speedups over existing systems, even for massive MoE models, without sacrificing reward convergence.