I am a master's student at the University of Chinese Academy of Sciences, advised by Prof. Xiangsheng Huang, and expect to graduate in June 2027. I am interested in how multimodal models learn representations, adapt to new tasks, and use experience to reason and act.
My recent work, FiberTune, studies how to preserve transferable visual structure during vision-language-action (VLA) fine-tuning. I also build robot systems and contribute to open-source software.
Research Interests
Multimodal representation learning and reasoning
Continual learning and adaptation
World modeling and embodied intelligence
Publications
FiberTune: Preserving Action-Fiber Visual Residuals in Vision-Language-Action Fine-Tuning
Haihao Lin, Xiangsheng Huang, Xiao Yang, Weibang Zhou, Yiqi Zhang, Bo Yang, Simin Zeng, Jiawei Yang, Zhengyang Wang, Jiahui Du.
Accepted at CoRL 2026. First and corresponding author.
[arXiv]
[PDF]
[Project]
[Code]
FiberTune preserves transferable visual structure during VLA fine-tuning, without adding inference-time overhead. Evaluation covers CALVIN, LIBERO, and a real-world SO-101 task. It improves CALVIN ABC→D five-step success from 61.4% to 72.1% and SO-101 task success from 72.7% to 78.1% over task-loss-only fine-tuning.
Honors and Awards
National Scholarship (twice); Jiangsu University President's Medal; Outstanding Undergraduate Thesis.
Three silver medals at ICPC Asia Regional Contests; highest-ranked bronze-medal team at the 2024 ICPC Asia-East Continent Final.
Two First Prize awards at the Lanqiao Cup National Finals, in C/C++ and Python (Python: 1st in Jiangsu, 9th nationally).
Engineering Projects
Commercial cleaning robot. Developed ROS2 systems for autonomous charging-dock return and docking, cliff detection and avoidance, and simulation-based testing before deployment. Also worked on coverage-path planning and speech interaction.
RGB-D grasping for logistics. Built an xArm-based tabletop grasping prototype on Jetson, including hand-eye calibration, suction and gripper integration, and Web API integration.
Open-Source Contributions
I was an early key contributor to
MaaAssistantArknights
(20k+ GitHub stars), where I developed C++ core modules, WPF GUI features, and a task-planning algorithm for mutual-exclusion constraints.
In the HoshinoBot ecosystem, I developed and maintained aichat and other plugins. These integrate models through OpenAI-compatible APIs and support multi-turn context, per-group configuration, persistent sessions, model switching, and voice replies. I also built non-invasive message-trigger conflict detection tools to reduce unintended plugin responses.
As captain of Jiangsu University's programming contest team, I helped establish its training and selection system. I independently set and reviewed all warm-up and main-contest problems and wrote their editorials for the 2022 and 2023 university contests, and also served as an on-site proctor and answered contestants' questions.
For Lanqiao Cloud Courses, I independently developed the Python preparation course for the 17th Lanqiao Cup, including its curriculum and algorithm content, and contributed problems to the online judge problem bank.
Education
M.S., Computer Science (Artificial Intelligence), University of Chinese Academy of Sciences, Sep. 2024–Jun. 2027 (expected).