Junze (Tony) Ye
Junze (Tony) Ye
I work on efficient post-training and evaluation of AI models under data or compute constraints. My research includes efficient on-policy distillation for LLM agents, inference-time fusion of generative experts, and adaptive physician querying for clinical benchmark labeling. From these, I accumulated experience with distributed model training, inference pipeline customization, and data quality control best practices.
Faculty Advisors
Research Interests
- Post-training
- Efficiency Optimization
- Adaptive Learning
- Applied Probability
Publications
Working Papers
Under journal review. Finalist for INFORMS HAS Student Paper Competition. ArXiv preprint: https://arxiv.org/abs/2512.19691 Code and data release: https://github.com/junzeye/validate-medcalc-labels
(Sept 2026 update: a revised version is accepted by NeurIPS 2026 Main) Accepted by ICML 2026 workshop: RL from World Feedback (RLxF). TL;DR: When finetuning LLM agents, spending teacher labels on broader student-context coverage can be more effective than spending it on longer or more heavily filtered teacher completions.
A poster version was presented at CS 329A's poster session on December 12, 2025.