Yinghui He (何映晖)
Hi everyone! I'm Yinghui He, a third-year PhD candidate at Princeton University Computer Science Department. I'm honored to be advised by Sanjeev Arora at Princeton Language and Intelligence (PLI).
I work on post-training and evaluation for LLMs and agents. I focus on how models gain the skills they're missing and how they learn from their own attempts. My work follows three lines:
- Skill-aware training: finding which skills a model lacks and aiming training at them (AdaptMI, STAT, Skill Entropy).
- Self-distillation: turning a model's own attempts into dense supervision. This covers reasoning (SD-Zero, Rethinking OPSD) and multi-turn agents that learn to recover from pivotal mistakes (PivotOPD).
- Benchmarking: measuring what LLMs still can't do well, including higher-order theory of mind (Hi-ToM), long procedural generation (LongProc), skill-switching in long-horizon reasoning (Skill Entropy) and mental-health safety in human–AI interaction (EmoAgent).
I finished my Bachelor's degree in Computer Science at the University of Michigan and Shanghai Jiao Tong University, where I had the honor to work with Rada Mihalcea (at the LIT Lab) and Wei Hu. I interned at NVIDIA Research in summer 2026, as a Research Scientist Intern.
