D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoders
Uses sparse autoencoders to decompose and steer diffusion transformers.
Interpretability
Demo
About Me
I am an incoming postdoctoral fellow at The Chinese University of Hong Kong (CUHK), where I will work with Prof. Weiyang Liu and Prof. Pheng-Ann Heng. I am also a Pivotal AI Safety Research Fellow, mentored by Dr. Peter Hase of Schmidt Sciences and Stanford University.
I completed my PhD in Electronic and Computer Engineering at HKUST in July 2026 as a Hong Kong PhD Fellowship awardee, advised by Prof. Xiaomeng Li. I also spent part of my PhD at the University of Cambridge as a visiting student, supervised by Prof. Adrian Weller and Prof. Weiyang Liu.
Before my PhD, I earned a Bachelor of Advanced Computing (Honours) at The Australian National University (ANU), specializing in Intelligent Systems. My honours research was supervised by Senior Prof. Amanda Barnard and Dr. Amanda Parker. I also held research internships at the National University of Singapore and the MIT CSAIL Computational Connectomics Group, where I worked with Prof. Lu Mi and Prof. Hao Wang.
I welcome research collaborations and would be happy to hear from you if our interests align.
Research Interests
- Interpretability: Concept bottleneck models, probabilistic modeling, mechanistic interpretability.
- LLM Safety: Faithfulness, hallucination mitigation, reliable multimodal reasoning.
- AI for Science and Healthcare: Biological modeling, medical image analysis, environmental modeling.
Selected Publications
* indicates equal contribution. Underlined names indicate equal advising.
Overview
Interpretable Machine Learning
Concept-Grounded Reasoning with Prompt-Driven Localization for Interpretable Structured Report Generation
Combines prompt-driven lesion localization with concept-grounded reasoning for interpretable medical report generation.
Concept-Based Unsupervised Domain Adaptation
Improves interpretability and robustness of Concept Bottleneck Models under domain shifts through adversarial training.
Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations
A unified framework for concept-based prediction, concept correction, and fine-grained interpretations based on conditional probabilities.
Energy-Based Conceptual Diffusion Model
Unifies concept-based generation, conditional interpretation, concept debugging, intervention, and imputation under a joint energy-based formulation.
Interpretability
Paper
Classification Models for Medical Data with Interpretative Rules
Recognizes positive SARS cases across numerical and categorical datasets, with explanatory rules generated for clearer interpretability.
Reasoning & Evaluation
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Reasoning
Faithful-First Reasoning, Planning, and Acting for Multimodal LLMs
Studies faithful-first reasoning, planning, and acting for multimodal large language models.
AI for Science & Healthcare
Multimodal fusion of complementary material representations for generalizable and interpretable property prediction
Combines composition, crystal structure, and radial distribution features for interpretable material property prediction and more stable out-of-distribution generalization.
AttTok: Marrying Attribute Tokens with Generative Pre-trained Vision-Language Models towards Medical Image Understanding
Integrates attribute tokens with generative pre-trained vision-language models for medical image understanding.
Machine Learning Models for Evaluating Biological Reactivity within Molecular Fingerprints of Dissolved Organic Matter over Time
Connects molecular composition and biological reactivity in reservoir data, yielding additional interpretability beyond predictive performance.
AI4Science
Interpretability
Exploring the Complexities of Dissolved Organic Matter Photochemistry from the Molecular Level by Using Machine Learning Approaches
Integrates irradiation experiments and provides insights into estuarine dissolved organic matter transformation.
Conference version: The 6th Xiamen Symposium on Marine Environmental Sciences, Best Poster Award.
Dynamic Data Augmentation via Monte-Carlo Tree Search for Prostate MRI Segmentation
Uses efficient dynamic augmentation to reduce overfitting in limited medical image segmentation settings.
Honors
- 2026 Pivotal AI Safety Research Fellowship
- 2026 ICML Gold Reviewer Award
- 2026 DAAD AInet Fellow
- 2023 - 2026 HKUST Academic Excellence Awards (Continuing Students)
- 2022 - 2026 Hong Kong PhD Fellowship
- 2023 - 2025 HKPFS Conference Travel Allowance
- 2022 - 2023 HKUST RedBird PhD Award
Service
Reviewer
- Conferences AAAI, ICLR, ICML, NeurIPS, CVPR, ECCV, EMNLP, MICCAI SAFER Workshop
- Journals TMLR, IEEE TPAMI, Knowledge-Based Systems (KBS)
Program Committee
- Conference ICONIP
Memberships
- Since 2021 IEEE Member
- Since 2021 Australian Computer Society Associate Member
- Since 2021 Asia Pacific Neural Network Society Member
Teaching
- Feb - Dec 2023 TA, ELEC 1200 A System View of Communications: from Signals to Packets, HKUST
- Dec 2019 - Jan 2020 TA, Statistics of Stochastic Process & Algorithms and Analysis, SJTU Summer School