Lin Li

pronounced like "Lynn Lee", and written "李淋" in Chinese.

prof_pic.jpg

Hello! I’m a postdoctoral researcher in the Oxford Applied and Theoretical Machine Learning (OATML) Group at the Department of Computer Science, University of Oxford, where I am advised by Prof. Yarin Gal.

I received my PhD from King’s College London, supervised by Prof. Michael Spratling, an MSc in Computing from Imperial College London under Prof. Wayne Luk, and a B.M. in Finance from Xiamen University, advised by Prof. Zheng Qiao.

My research interests include

  • Hallucination in (multimodal) LLMs
  • LLM Jailbreaking, Safety Alignment, and Adversarial ML
  • AI Society and LLM Agent
  • AI+ Applications: healthcare, robotics, business

I’m always open to collaboration and connection, so feel free to get in touch! I’m also looking for self-motivated students to explore exciting research directions together — if you’re interested, reach out and let’s grab a coffee (to update my selfie here 😄).

News:

03/01/25 A new co-authored article, titled “Reducing Large Language Model Safety Risks in Women’s Health using Semantic Entropy”, is available at arXiv now.
02/17/25 Invited to review for ICML 2025 and NeurIPS 2025.
02/15/25 My first-authored work, Robust shortcut and disordered robustness: Improving adversarial training through adaptive smoothing, is online now at the Pattern Recognition (PR).
02/09/25 Serve as the Program Committee of the 2nd MEIS Workshop @CVPR2025.
02/09/25 Invited to review for the journals IEEE T-PAMI, IEEE T-IFS, IEEE T-DSC, and Pattern Recognition.
01/21/25 I gave an invited talk, titled “Prompting Vision-Language Models for Accuracy and Robustness”, at the Computational Health Informatics (CHI) Lab in the Department of Engineering Science at Oxford.
11/21/24 A new co-authored article, titled “Artificial Intelligence for Biomedical Video Generation”, is online at Arxiv.
09/24/24 My first-authored (equal contribution) work, Advancing robots with greater dynamic dexterity: A large-scale multi-view and multi-modal dataset of human-human throw&catch of arbitrary objects, is online now at the International Journal of Robotics Research (IJRR).
08/24/24 A first-authored work, AROID: Improving Adversarial Robustness through Online Instance-wise Data Augmentation, is accepted by the International Journal of Computer Vision (IJCV).
07/20/24 Invited to serve as Program Committee for AAAI 2025. What a fancy name of reviewer!
06/07/24 A first-authored work, Advancing Robots with Greater Dynamic Dexterity: A Large-Scale Multi-View and Multi-Modal Dataset of Human-Human Throw&Catch of Arbitrary Objects, is accepted by the International Journal of Robotics Research (IJRR).
05/02/24 Our work, OODRobustBench: a benchmark and large-scale analysis of adversarial robustness under distribution shift, is accepted by ICML 2024. Be careful! your adversarially robust model is probably much less robust under distribution shifts.
04/26/24 I will attend the conference VALSE2024 at Chongqing, China from 5th to 7th May and present my work of APT at the poster session (the poster ID is 331) on 6th May. Look forward to chat!
03/27/24 I was invited by AI Time to give a talk of our recent CVPR2024 publication: One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models.
03/04/24 Our work, OODRobustBench: benchmarking and analyzing adversarial robustness under distribution shift, is accepted by ICLR 2024 Workshop Data-centric Machine Learning Research (DMLR)
02/27/24 One work, One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models, is accepted by CVPR2024!
12/23/23 I am invited to serve as reviewer for Internation Conference on Machine Learning (ICML) 2024.
12/22/23 I join the program committee of Workshop on Wearable Intelligence for Healthcare Robotics (WIHR): from Brain Activity to Body Movements at 2024 IEEE International Conference on Robotics and Automation (ICRA) in PACIFICO Yokohama, Japan.
11/23/23 I am invited to serve as reviewer for the journal of IEEE Transactions on Dependable and Secure Computing.
10/24/23 A new work of OODRobustBench: benchmarking and analyzing adversarial robustness under distribution shift is pre-printed on Arxiv now.
10/08/23 A new co-authored work of VisionFM: a Multi-Modal Multi-Task Vision Foundation Model for Generalist Ophthalmic Artificial Intelligence is pre-printed on Arxiv now.
09/24/23 Our review of Large AI Models, a.k.a. foundation models, in healthcare has been accepted by IEEE Journal of Biomedical and Health Informatics (JBHI).
06/15/23 I will present my research about adversarial robustness at the event King’s College London, Department of Informatics Research Showcase on 20th June, feel free to come and talk.
06/13/23 our work of improving adversarial robustness through online instance-wise automated data augmentation is pre-printed on Arxiv now.
05/07/23 I will give a talk about our ICLR2023 paper “DA alone can improve AT” on 30 May as invited by AI Time.
04/10/23 I will serve as a reviewer for NeurIPS 2023.
03/28/23 Our work of Instance-adaptive Smoothness Enhanced AT (ISEAT) is pre-printed on Arxiv now.
03/22/23 Our review of Large AI (language, vision, multi-modalities) Models a.k.a. foundation models in Health Informatics is pre-printed on Arxiv now.
01/21/23 Our work of data augmentation for adversarial training, IDBH, was accepted at ICLR2023.
11/30/22 Our work of Understanding and combating robust overfitting via input loss landscape analysis and regularization was accepted at Pattern Recognition.

Selected Publications:

  1. AROID: Improving Adversarial Robustness Through Online Instance-Wise Data Augmentation
    Lin Li†, Jianing Qiu, and Michael Spratling
    International Journal of Computer Vision (IJCV), 2024
  2. One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models
    Lin Li*, Haoyan Guan*, Jianing Qiu, and Michael Spratling
    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024
  3. OODRobustBench: a benchmark and large-scale analysis of adversarial robustness under distribution shift
    Lin Li, Yifei Wang, Chawin Sitawarin, and Michael Spratling
    International Conference on Machine Learning (ICML) and ICLRW-DMLR, 2024
    Benchmark
  4. Advancing Robots with Greater Dynamic Dexterity: A Large-Scale Multi-View and Multi-Modal Dataset of Human-Human Throw&Catch of Arbitrary Objects
    Lipeng Chen*†, Jianing Qiu*, Lin Li*, Xi Luo, Guoyi Chi, and Yu Zheng
    International Journal of Robotics Research (IJRR), 2024
    Web
  5. Data Augmentation Alone Can Improve Adversarial Training
    Lin Li, and Michael Spratling
    International Conference on Learning Representations (ICLR), 2023
  6. Large AI Models in Health Informatics: Applications, Challenges, and the Future
    Jianing Qiu, Lin Li, Jiankai Sun, Jiachuan Peng, Peilun Shi, Ruiyang Zhang, Yinzhao Dong, Kyle Lam, Frank P.-W. Lo, Bo Xiao, and 3 more authors
    IEEE Journal of Biomedical and Health Informatics (J-BHI), 2023