Scholar
Difan Zou
Google Scholar ID: Cp4fcTQAAAAJ
The University of Hong Kong
Machine Learning
Deep Learning
Optimization
Stochastic Algorithms
Signal Processing
Follow
Homepage
↗
Google Scholar
↗
Citations & Impact
All-time
Citations
5,440
H-index
28
i10-index
61
Publications
20
Co-authors
18
list available
Contact
Email
dzou@cs.hku.hk
Twitter
Open ↗
GitHub
Open ↗
Publications
62 items
Understanding the Weight Averaging Mechanism in LLM Training for Post-Training Quantization
2026
Cited
0
Less Sycophancy, Stronger Refusal? Lessons for AI Safety from Mechanistic Interpretability
2026
Cited
0
Beyond Token Scale: Chunk-Level Sparse Autoencoders for Reliable Semantic Feature Discovery
2026
Cited
0
Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders
2026
Cited
0
CoRe: Co-Evolving Reward Models for Mitigating Latent Reward Hacking in Video Diffusion Models
2026
Cited
0
From Static to Dynamic: On-Policy Distillation from Image to Video Diffusion Models
2026
Cited
0
SkillRubric: Co-Evolving Actor Guidance and Evaluator Rubrics for Multimodal Agents
2026
Cited
0
Omni-Streaming Thinking
2026
Cited
0
Load more
Resume
Academic Achievements
Multiple papers accepted at top-tier conferences including NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, COLT, and CoLLAs
2025: 7 NeurIPS, 3 ICML, 4 ICLR, 1 ACL, 1 EMNLP, 1 CVPR
2024: 5 NeurIPS, 4 ICML, 3 ICLR, 1 COLT, 1 CoLLAs; one paper on faster diffusion inference won Best Paper Award (oral) at ICML SPIGM workshop
2023: 1 COLT (implicit bias of batch normalization), 1 ICLR (generalization gap between Adam and GD), served as Area Chair for NeurIPS 2023
2022: 2 NeurIPS papers (on multi-pass SGD generalization and pretraining-finetuning limits in linear regression with distribution shift)
Background
Assistant Professor, Department of Computer Science, The University of Hong Kong
Affiliated with the School of Computing and Data Science and The HKU Musketeers Foundation Institute of Data Science
Research interests include machine learning, stochastic optimization, and graph learning
Special focus on theoretical/empirical understanding (or physics) of deep learning, especially foundation models
Developing AI/ML methods for practical problems in signal processing, intelligent transportation, and mathematical problems
Co-authors
17 total
Quanquan Gu
Associate Professor of Computer Science, UCLA
Yuan Cao
The University of Hong Kong
Pan Xu
Duke University
Jingfeng Wu
University of California, Berkeley
Vladimir Braverman
Professor of Computer Science, Johns Hopkins University; Google Research; Adjunct Professor, Rice U.
Sham M Kakade
Harvard University
Dongruo Zhou
Indiana University Bloomington
Jinghui Chen
Assistant Professor of Information Sciences and Technology, Penn State University