My research focuses on building scalable models to achieve 99% accuracy on scalable test datasets, applicable to NLP, CV, VLA, LLM and MLLM.

My current observation is that generalized re-label is all you need. This can manifest as re-labeling at the data level, the prompt level, or other dimensions.