Yingbin Liang

Discrete diffusion models can now generate samples in polylogarithmic time, thanks to a novel Gibbs-based corrector that avoids costly retraining.

LLMs can now reliably follow complex, hierarchical instructions thanks to a new constrained RL framework that treats system prompts as strict algorithmic boundaries.

Adam's faster convergence isn't just empirical luck: its second-moment normalization provably yields sharper tails in high-probability convergence guarantees compared to SGD.

RLVR's success in long-horizon reasoning hinges on a smooth difficulty spectrum, where mastering easier sub-problems unlocks the ability to tackle harder ones, avoiding frustrating grokking plateaus.