NVIDIA
Current May 2026 – PresentApplied Deep Learning Research Intern · Mentor: Wei Ping
Token-efficient reinforcement learning for LLM reasoning — getting more capability per training and inference token.
Highlights
Recent milestones, newest first.
-
May 2026
Joined NVIDIA — Applied Deep Learning Research Intern
Working on token-efficient reinforcement learning for LLM reasoning with Wei Ping.
-
2026
NorMuon selected as an ICML 2026 Spotlight Top 2.2%
Our optimizer set new records on the modded-nanogpt speedrun leaderboard and was adopted in Andrej Karpathy's nanochat.
-
2025
Interned at Microsoft — two papers accepted at ICML 2026
The work led to NorMuon and Shuffle the Context (long-context adaptation via RoPE-perturbed self-distillation), both published at ICML 2026.
-
2024
Interned at Microsoft — released compact Phi MoE models
SlimMoE (COLM 2025) shipped as Phi-mini-MoE and Phi-tiny-MoE on Hugging Face — 6M+ total downloads.
-
2024
Two RLHF papers accepted at NeurIPS 2024
Adaptive preference scaling for preference optimization, and robust reward learning from corrupted human feedback.
Experience
Research internships at industry labs.
Microsoft
June 2025 – Aug 2025Research Intern · Mentor: Chen Liang
Long-context adaptation of LLMs — RoPE-perturbed self-distillation, published as Shuffle the Context at ICML 2026.
Microsoft
May 2024 – Aug 2024Research Intern · Mentor: Chen Liang
Structured compression and distillation of large MoE models — shipped as Phi-mini-MoE and Phi-tiny-MoE (6M+ downloads), published as SlimMoE at COLM 2025.
Publications
* denotes equal contribution.
Language models
-
ICML 2026Spotlight · Top 2.2%
NorMuon: Making Muon More Efficient and Scalable
New records on the modded-nanogpt speedrun; adopted in Andrej Karpathy's nanochat.
-
ICML 2026
Shuffle the Context: RoPE-Perturbed Self-Distillation for Long-Context Adaptation
-
ICLR 2026
COSMOS: A Hybrid Adaptive Optimizer for Efficient Training of Large Language Models
-
COLM 2025
SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation
Released as Microsoft's open Phi-mini-MoE and Phi-tiny-MoE — 6M+ downloads.
-
arXiv 2025
LLMs Can Generate a Better Answer by Aggregating Their Own Responses
-
IEEE TMC 2025
Mitigating Tail Latency for On-Device Inference with Load-Balanced Heterogeneous Models
-
NeurIPS 2024
Adaptive Preference Scaling for Reinforcement Learning with Human Feedback
-
NeurIPS 2024
Robust Reinforcement Learning from Corrupted Human Feedback
Point processes & other
-
ICML 2024
Beyond Point Prediction: Score Matching-based Pseudolikelihood Estimation of Neural Marked Spatio-Temporal Point Process
-
ICML 2023
SMURF-THP: Score Matching-based Uncertainty Quantification for Transformer Hawkes Process
-
ICDE 2023
Efficient Deep Ensemble Inference via Query Difficulty-dependent Task Scheduling
-
WWW 2023
CoTel: Ontology-Neural Co-Enhanced Text Labeling
-
ICML 2020
Transformer Hawkes Process
A standard method for neural temporal point processes — co-developed as an undergraduate.
-
Software
PRIMAL: A Linear Programming-based Sparse Learning Library in R and Python
Education
Georgia Institute of Technology
Aug 2023 – Present · Atlanta, GAPh.D. in Machine Learning
Advisor: Prof. Tuo Zhao
University of Science and Technology of China
Rank 1 / 56 Sept 2020 – June 2023 · Hefei, ChinaM.S. in Data Science
Advisor: Prof. Lan Zhang
University of Science and Technology of China
Admitted at 14 Sept 2016 – June 2020 · Hefei, ChinaB.S. in Mathematics and Applied Mathematics · School of the Gifted Young
Entered university at 14 through USTC's selective early-entrance program; graduated at 18.