Yan-Shuo Liang 「梁宴硕」

I am currently a PhD candidate in the Department of Computer Science and Technology at Nanjing University, supervised by Prof. Wu-Jun Li (李武军).

I received my B.Sc. degree from the Department of Mathematics at Nanjing University in 2020. In the same year, I was admitted to study for a PhD degree without entrance examination. I will complete my PhD in 2026 and am currently actively seeking full-time or internship opportunities in both industry and academia.

CV / Github / Google Scholar / DBLP / Email: liangys@smail.nju.edu.cn

Biography

2024.03-2026.03: PhD candidate, Department of Computer Science and Technology, Nanjing University
2025.05-2025.09: Internship at ByteDance, Hangzhou
2020.09-2024.03: PhD student, Department of Computer Science and Technology, Nanjing University
2016.09-2020.06: B.S., Department of Mathematics, Nanjing University

Research Interests

Continual / Incremental Learning: Minimize interference while learning new tasks and maintain stable long-term learning ability.
Parameter-Efficient Fine-Tuning: Freeze backbone of large models, optimize lightweight inserted modules for efficient training under resource constraints.

Internship Experience

ByteDance, Hangzhou (May 2025 – Sep 2025)

Department: Data-AML – Applied Algorithms Team
Position: Large Model Algorithm Intern (Large Systems & Compute) – Soaring Star Talent Program

Publications (First Author)

Gated Integration of Low-Rank Adaptation for Continual Learning of Language Models

Yan-Shuo Liang, Wu-Jun Li

Conference on Neural Information Processing Systems (NeurIPS), 2025, CCF-A First Author

paper code bibtex

Proposed Gated Integration of Low-Rank Adaptation (GainLoRA), introducing a gating network on top of InfLoRA. Uses a mixture-of-experts paradigm for multi-task incremental learning.

InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning

Yan-Shuo Liang, Wu-Jun Li

Computer Vision and Pattern Recognition Conference (CVPR), 2024, CCF-A First Author

paper code bibtex

We propose a new PEFT method, called InfLoRA, for continual learning. InfLoRA injects a small number of parameters to reparameterize the pre-trained weights and shows that fine-tuning these injected parameters is equivalent to fine-tuning the pre-trained weights within a subspace. Furthermore, InfLoRA designs this subspace to eliminate the interference of the new task on the old tasks.

Loss Decoupling for Task-Agnostic Continual Learning

Yan-Shuo Liang, Wu-Jun Li

Annual Conference on Neural Information Processing Systems (NeurIPS), 2023, CCF-A First Author

paper code bibtex

We propose a simple yet effective method, which separates the two objectives for the new task by decoupling the loss of the new task, providing a way to obtain a better trade-off between stability and plasticity than those methods with coupled loss.

Adaptive Plasticity Improvement for Continual Learning

Yan-Shuo Liang, Wu-Jun Li

Computer Vision and Pattern Recognition Conference (CVPR), 2023, CCF-A First Author

paper code bibtex

We propose a continual learning framework, Adaptive Plasticity Improvement (API), which achieves the decoupling of new and old task learning through a parameter dynamic expansion mechanism. Subsequent work extends the API framework to pre-trained models and large language models.

Awards & Honors

博士生校长特别奖学金. Presidential Special Scholarship, Nanjing University, 2020
华为奖学金. Huawei Scholarship, Nanjing University, 2023
南京大学优秀研究生标兵. Outstanding Graduate Student of Nanjing University, 2024
博士生国家奖学金. National Scholarship for PhD Students, Nanjing University, 2024

Teaching Assistant

Data Structure (Undergraduate, Autumn 2020)
Data Structure (Undergraduate, Spring 2021)

Contact

Email: liangys [at] smail [dot] nju [dot] edu [dot] cn

Nanjing University Xianlin Campus, 163 Xianlin Avenue, Qixia District, Nanjing, Jiangsu 210023, China