ABOUT ME
I am an undergraduate in the Yao Class at the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University.
I am currently a research intern at MIT Han Lab, advised by Prof. Song Han. I also work with Prof. Wei Xu at Tsinghua on distributed systems for large language model training.
My research focuses on machine learning systems: making large-model training and inference more efficient and reliable, and improving how language models learn to reason.
Introduced future-aware on-policy distillation for block diffusion language models. The method corrects an autoregressive teacher's targets using visible future context within each block, aligning supervision with the information available to the diffusion student.
Improves model performance while making distillation more efficient.
Developed Semantic Masked Expert Policy Optimization (SMEPO) for expert-guided reinforcement learning. The method preserves useful reasoning structure while masking information that could let a model copy its way to a reward.
Evaluated on mathematical reasoning, code, and agent tasks.
Designed and implemented a runtime that dispatches ready microbatch work in a Megatron-based training framework. Existing schedules provide priorities, while actual dependencies determine what can execute.
Dynamic pipeline execution that adapts to computation and communication variability.
Worked on fast failure recovery for large language model training, restoring training state with minimal additional state-management overhead.
Optimized pipeline execution order and dependencies to increase computation–communication overlap and reduce pipeline bubbles in Megatron-based training.
Contributed to a diffusion framework for coordinated interactions among humans, hands, and objects, including model development, training, and data processing.
Designed microstate-based representations and deep learning models for EEG signal modeling and analysis.
Implemented FunctionCoverRelation and Threshold Relation checks, together with their online-checking versions. Ran experiments to collect relations and analyze inferred training invariants.
Oct 2025 – Present
Research Intern · Advisor: Prof. Song Han
Reinforcement learning and on-policy distillation for large language models.
Aug 2024 – Present
Research with Prof. Wei Xu
Distributed training, pipeline parallelism, and fast failure recovery.
Jul – Sep 2025
Research with Prof. Ryan Huang
Automated sanity checks for deep learning training.
Feb – Aug 2024
Research with Prof. Li Yi
Generative modeling for multi-body human–object interactions.
Institute for Interdisciplinary Information Sciences
Tsinghua University
GPA: 3.9 / 4.0 · Rank: 8 / 90
Introduction to Computer Systems2025, 2026
Calculus2024, 2025
President, 2nd Student Science and Technology Association, IIIS, Tsinghua University
第二届交叉信息院学生科协主席
Yao Award姚奖
Yanbao Scholarship燕宝奖学金 · Comprehensive Scholarship
Zheng Geru Scholarship郑格如奖学金 · Comprehensive Scholarship
Xuetang Scholarship清华大学学堂奖学金
Freshman Scholarship新生奖学金