Preprint, 2026
Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control
Entropy-curve control for stable long-horizon RLVR training.
Personal Information
I am a PhD candidate in the Department of Computer Science at Purdue University, advised by Ruqi Zhang. I completed my BE in computer science and technology at Tianjin University, advised by Changqing Zhang.
Research Interests
My research develops statistical foundations for stable and efficient post-training frameworks. I currently focus on improving the performance boundary of LLM RL, including RL entropy control, and agentic MoE RL. More broadly, I work on test-time scaling, preference alignment, multimodal LLM safety, and Bayesian/statistical methods for robust ML.
Featured First-Author Papers
EMNLP Main, 2025
Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner
Speculative decoding reframed as efficient reward-guided alignment.
COLM, 2025
Cascade Reward Sampling for Efficient Decoding-Time Alignment
Segment-level rejection sampling for faster aligned generation.
TMLR, 2025
Making Reliable and Flexible Decisions in Long-tailed Classification
Decision-theoretic utilities for reliable long-tailed predictions.
ICLR, 2024
Entropy-MCMC: Sampling from Flat Basins with Ease
A flatness-aware sampling method for Bayesian deep learning.
CVPR, 2022
Trustworthy Long-tailed Classifications
Uncertainty-aware expert engagement for long-tailed recognition.
WWW Oral, 2022
Graph Communal Contrastive Learning
Community-aware contrastive learning for graph representation.