Ph.D. Candidate in Computer Science specializing in AI/ML
Hong Kong, HK.
About
Zhan Zhuang is a Ph.D. Candidate in Computer Science at CityUHK, specializing in post-training and transfer learning for foundation models, with an expected graduation in June 2026. He boasts an extensive publication record in top-tier AI/ML venues including ICLR, ICML, NeurIPS, and ICCV, complemented by industry experience from ByteDance. His expertise in parameter-efficient updates, automated self-improvement, and long-context memory mechanisms positions him to develop robust, generalizable, and self-improving AI systems.
Work
ByteDance Seed Research
|
Research Intern
Shenzhen, Guangdong, China
→
Summary
Conducted hands-on industry research at ByteDance Seed Research, contributing to advanced projects in foundation models and self-improving learning systems.
Highlights
Developed and implemented parameter-efficient update strategies for foundation models, enabling frequent, low-cost model refreshes while preserving prior capabilities.
Contributed to the design and evaluation of automated self-improvement mechanisms, utilizing curriculum/data synthesis and self-evaluation for robust model performance.
Explored and integrated long-context and memory architectures, significantly improving task completion and generalization for complex AI systems.
Applied hands-on industry experience to optimize large-scale AI systems, ensuring deployment feedback loops for continuous improvement and robust generalization.
Education
City University of Hong Kong (CityUHK)
Hong Kong, Hong Kong SAR, Hong Kong
→
Ph.D.Computer Science
Southern University of Science and Technology (SUSTech)
Shenzhen, Guangdong, China
→
B.Eng.Computer Science and Technology
Awards
Outstanding Academic Performance Award
CityUHK
Recognized for exceptional academic achievement and contributions during Ph.D. studies.
Outstanding Graduate Award
SUSTech
Awarded for distinguished academic performance and overall contributions as a graduating student.
Top 10 Graduate Award at Shude College
SUSTech
Acknowledged as one of the top 10 graduates from Shude College, demonstrating academic excellence.
Excellent Student Cadre
SUSTech
Recognized for outstanding leadership and service in student organizations.
Excellent League Member
SUSTech
Awarded for exemplary commitment and active participation within the student league.
Publications
Method and device for acquiring information of close contact person, server and storage medium
Granted Patent (CN112866991B)
Invented and patented a method and device for acquiring close contact information, enhancing public health initiatives.
Data processing method and device based on block chain, equipment and storage medium
Granted Patent (CN112635061B)
Developed and patented a novel data processing method and device leveraging blockchain technology for secure and efficient information management.
HAD: Heterogeneity-Aware Distillation for Lifelong Heterogeneous Learning
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Authored research on heterogeneity-aware distillation, advancing lifelong heterogeneous learning in computer vision.
MoSA: Mosaic Shared Adaptation of Large Language Models
Thirteenth International Conference on Learning Representations (ICLR)
Co-authored work on mosaic shared adaptation, improving parameter-efficient fine-tuning for large language models.
PLAN: Proactive Low-Rank Allocation for Continual Learning
International Conference on Computer Vision (ICCV)
Introduced PLAN, a proactive low-rank allocation strategy to optimize continual learning performance.
HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models
Thirteenth International Conference on Learning Representations (ICLR)
Presented HiRA, an oral presentation (top 1.8%) on parameter-efficient Hadamard high-rank adaptation for LLMs, enhancing model efficiency.
Moonshine: Distilling Game Content Generators into Steerable Generative Models
Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI)
Delivered an oral presentation on Moonshine, a method for distilling game content generators into steerable generative models.
Come Together, But Not Right Now: A Progressive Strategy to Boost Low-Rank Adaptation
Forty-Second International Conference on Machine Learning (ICML)
Proposed a progressive strategy to significantly boost low-rank adaptation in machine learning models.
MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition
Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS)
Developed MoPFormer, a novel transformer architecture for enhancing wearable-sensor activity recognition.
Gradual Domain Adaptation via Gradient Flow
Twelfth International Conference on Learning Representations (ICLR)
Published research on gradual domain adaptation using gradient flow, recognized as a Spotlight paper (top 5%) for its innovative approach.
Extend Wave Function Collapse Algorithm to Large-Scale Content Generation
IEEE Conference on Games
Awarded 'Best Paper' (top 5%) for extending the Wave Function Collapse algorithm to enable large-scale content generation.
Linux, Android, Unity, Arduino, Web Development, STM32, Jetson Nano, Hololens.
Machine Learning & AI
Foundation Models, Transfer Learning, Continual Learning, Parameter-Efficient Fine-Tuning (PEFT), Large Language Models (LLM), Computer Vision (CV), Neural Information Processing Systems, Deep Learning, Generative Models, Reinforcement Learning.
Research Methodologies
Algorithmic Screening, Data Synthesis, Self-Evaluation, High-Performance Computing (HPC), Model Optimization, Deployment Feedback Loops, Strategic Prioritization.