Parameter-Efficient LLM Adaptation
Developing low-rank and progressive adaptation strategies for large language models.
PEFT
LoRA
Efficiency
Machine Learning + Large Language Models
Efficient, reliable, and adaptable learning systems.
I am a Master's candidate in Computer Science at the Southern University of Science and Technology (SUSTech) in Shenzhen, China. My research spans machine learning and large language models, with a focus on parameter-efficient adaptation, continual learning, rigorous evaluation, and reproducible research.
Research Focus
My work explores efficient adaptation for large language models, continual learning, evaluation and benchmarks, and vision-language model alignment.
Continual Learning
Proactive allocation methods that preserve performance across evolving tasks.
Low-rank allocation
Stability
Transfer
Vision-Language Models
Normalizing soft prompts and aligning multimodal representations.
Soft prompts
VLMs
Alignment
Evaluation & Benchmarks
Probing model priors and reasoning with targeted diagnostics and benchmarks.
Diagnostics
Benchmarks
Robustness
Selected Publications
Corresponding author is marked with a dagger in the CV.
-
MoSA: Mosaic Shared Adaptation of Large Language ModelsManuscript
-
One-Token Verification for Reasoning LLMs, Anytime, AnywhereManuscript
-
Image Sorting as a Probe of Physical Prior Encoding in Vision ModelsPreprint
-
PLAN: Proactive Low-Rank Allocation for Continual LearningICCV 2025
-
Come Together, But Not Right Now: A Progressive Strategy to Boost Low-Rank AdaptationICML 2025
-
Nemesis: Normalizing the Soft-prompt Vectors of Vision-Language ModelsICLR 2024 (Spotlight)
Blog
Notes on experiments, benchmarks, and the research process.
Designing LoRA schedules that scale with tasks
A short guide to staging low-rank adapters across a growing task list without leaking evaluation signal.
Building a lightweight evaluation harness for reasoning LLMs
How I structure small, scriptable evaluations for one-token verification and chain-of-thought ablations.
Continual learning notes: proactive low-rank allocation
Lessons learned when moving from static ranks to a proactive allocator across a task stream.
Reproducible research notes: minimal notebooks, maximal signal
A simple workflow for capturing experiments without drowning in logs, branches, and reruns.
No posts match that search.
Research Projects
Current research themes connected to ongoing papers.
MoSA for LLM Adaptation
Mosaic shared adaptation strategies that improve efficiency without sacrificing quality.
PLAN for Continual Learning
Proactive low-rank allocation to sustain performance over long task sequences.
Nemesis for VLM Prompts
Normalizing soft prompts to stabilize and align vision-language models.
Teaching Assistant
Recent teaching assistant roles at SUSTech.
CSE5001 Advanced Artificial Intelligence
Teaching Assistant, Spring 2024.
CS201 Discrete Mathematics
Teaching Assistant, Spring 2021 and 2024; Fall 2021 and 2023.
CS208 Algorithm Design and Analysis
Teaching Assistant, Spring 2022.
Academic Service
- Reviewer: NeurIPS 2024 Workshops
- Reviewer: CVPR 2025
Honors & Awards
- First-class Scholarship for Postgraduate Students (Sep 2025)
- First-class Scholarship for Postgraduate Students (Sep 2024)
- Third Place, SUSTech Collegiate Programming Contest (Sep 2021)
- Excellent Student, Zhixin College, SUSTech (Apr 2020)
Contact
I am open to research collaborations and internships in machine learning and large language models. Email me with a short note about your interests.
wangxiequn@gmail.com
Based in Shenzhen, China.