2026
arXiv preprint
Visualizing Distribution Coverage in Generative Diffusion Models
Pass@k reveals how distribution-matching distillation trades broad coverage of valid outputs for stronger single-sample quality in image and video generation.
CV
Research Focus
Long-horizon generation, representation learning, and reasoning for generative models. Efficient training and inference; privacy and trustworthy machine learning.
Education
Carnegie Mellon University · August 2026–September 2030 (expected)
Ph.D. student in Societal Computing, School of Computer Science
Software and Societal Systems Department
Advisors: Zhiwei Steven Wu, Niloofar Mireshghallah, and Andrew Ilyas
Rice University · September 2025–June 2026
Ph.D. studies in Computer Science
Research with Chen Wei
Shanghai Jiao Tong University · September 2022–March 2025
M.Sc. in Computer Science
Shanghai Jiao Tong University · September 2018–June 2022
B.S. in Computer Science and Technology
Publications
Google Scholar · * Equal contribution
2026
arXiv preprint
Scaling Video Generation for Reasoning: At What Cost?
A controlled Rubik's Cube benchmark reveals the compute cost of hidden-state reasoning and why lower generation loss does not guarantee better reasoning.
2026
arXiv preprint
Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers
Cross-view class-token alignment improves ImageNet linear-probe accuracy by 9.4 percentage points while preserving comparable generation quality.
2026
arXiv preprint
Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation
PACC learns compact memory from a frozen video generator, retaining earlier events and improving consistency in long video generation.
2026
arXiv preprint
Taming Outlier Tokens in Diffusion Transformers
Dual-Stage Registers reduce harmful outlier tokens in both vision encoders and diffusion transformers, improving generation quality across ImageNet and text-to-image settings.
2026
TMLR 2026 (Featured Certification)
Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding
Selective Jacobi decoding speeds up autoregressive normalizing-flow inference by up to 4.7 times while preserving generation quality, without additional training.
2026
KDD 2026
Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks
We identify a corruption stage during few-shot diffusion fine-tuning and use Bayesian neural networks to improve fidelity, quality, and diversity without extra inference cost.
2025
ICCV 2025
Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment
Adaptive layer-wise gradient alignment exposes weaknesses in federated-learning defenses by preserving backdoor effectiveness while making malicious updates harder to detect.
2025
NeurIPS 2025
Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM
Exact unlearning can still leak removed examples when pre- and post-unlearning models are exposed; our guided attack doubles extraction success in some settings.
2025
TPDP 2025
Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis
A lightweight learned membership inference attack reveals privacy leakage in diffusion-based tabular synthesis and won all four tracks of the MIDST Challenge.
2025
ICML 2025
Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models
FineXtract uses the distribution shift between pretrained and personalized diffusion models to recover about 20% of fine-tuning data from real-world checkpoints.
2025
ICLR 2025 (Spotlight)
Targeted Attack Improves Protection against Unauthorized Diffusion Customization
Carefully selected targeted attacks provide stronger protection against unauthorized diffusion customization than untargeted adversarial watermarks.
2024
CVPR 2024
CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion
CGI-DM visualizes conceptual differences between pretrained and fine-tuned diffusion models to provide evidence for copyright authentication.
2024
ICLR 2024
Toward Effective Protection against Diffusion-Based Mimicry through Score Distillation
Score distillation makes protective image perturbations faster and more memory-efficient while strengthening defenses against unauthorized diffusion-based mimicry.
2023
ICML 2023 (Oral)
Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples
AdvDM uses adversarial examples as protective watermarks that prevent diffusion models from learning an artist's style from unauthorized images.
2023
arXiv preprint
Mist: Towards Improved Adversarial Examples for Diffusion Models
Mist improves the transferability and robustness of adversarial watermarks designed to protect artwork from unauthorized diffusion-model imitation.
Awards
MIDST Challenge, SaTML 2025 — Winning team, all four tracks.
Membership inference on diffusion-based synthetic tabular data.
Challenge results · Related work: TPDP 2025 workshop
Open Source Impact
Mist — Founder and main developer.
1,000+ combined GitHub stars across Mist and Mist-v2; featured in The Verge.
Social engagement: 16K reposts and 20K likes.
Academic Service
Reviewer: ECCV, ICML, NeurIPS, ICLR.