Xuefeng Li

UMoE transforms underperforming expert pools into high-performing domain-specific models, achieving up to 6.0 points improvement on key benchmarks without increasing computational costs.

AI coding agents that ace small code modernization tasks fall apart on larger repositories, with pass rates plummeting from 91% to 15% when scaling from 10K to 50K lines of code.

A class-dependent reward Transformer model beats NASA's solar flare predictions, but only when fed the right "knowledge-informed" features.