publications

2025

2025

1. preprint

   Metascale: Test-time scaling with evolving meta-thoughts

   Qin Liu, Wenxuan Zhou, Nan Xu, James Y Huang, Fei Wang, Sheng Zhang, Hoifung Poon, and Muhao Chen

   In arXiv preprint 2025
2. preprint

   VLM-Guard: Safeguarding Vision-Language Models via Fulfilling Safety Alignment Gap

   Qin Liu, Fei Wang, Chaowei Xiao, and Muhao Chen

   In arXiv preprint 2025
3. preprint

   A survey on mechanistic interpretability for multi-modal foundation models

   Zihao Lin, Samyadeep Basu, Mohammad Beigi, Varun Manjunatha, Ryan A Rossi, Zichao Wang, Yufan Zhou, Sriram Balasubramanian, Arman Zarei, Keivan Rezaei, and others

   In arXiv preprint 2025
4. EMNLP

   QA-LIGN: Aligning LLMs through Constitutionally Decomposed QA

   Jacob Dineen, Aswin RRV,  Qin Liu, Zhikun Xu, Xiao Ye, Ming Shen, Zhaonan Li, Shijie Lu, Chitta Baral, Muhao Chen, and others

   In Findings of EMNLP 2025
5. EMNLP

   Familiarity-aware evidence compression for retrieval augmented generation

   Dongwon Jung,  Qin Liu, Tenghao Huang, Ben Zhou, and Muhao Chen

   In Findings of EMNLP 2025
6. ACL

   SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment

   Qin Liu, Fei Wang, Chaowei Xiao, and Muhao Chen

   In Proceedings of ACL 2025
7. ACL

   Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models

   Qin Liu, Chao Shang, Ling Liu, Nikolaos Pappas, Jie Ma, Neha Anna John, Srikanth Doss, Lluis Marquez, Miguel Ballesteros, and Yassine Benajiba

   In Findings of ACL 2025
8. NAACL

   Test-time Backdoor Mitigation for Black-box Large Language Models with Defensive Demonstrations

   Wenjie Mo, Jiashu Xu,  Qin Liu, Jiongxiao Wang, Jun Yan, Chaowei Xiao, and Muhao Chen

   In Findings of NAACL 2025
9. NAACL Demo

   MetaScientist: A Human-AI Synergistic Framework for Automated Mechanical Metamaterial Design

   Jingyuan Qi, Zian Jia, Minqian Liu, Wangzhi Zhan, Junkai Zhang, Xiaofei Wen, Jingru Gan, Jianpeng Chen,  Qin Liu, Mingyu Derek Ma, and others

   In Proceedings of NAACL: System Demonstrations 2025
10. ICLR

    MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding

    Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li,  Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Lorena Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen

    In The Thirteenth International Conference on Learning Representations 2025

    HTML

2024

2024

1. preprint

   Unraveling Cross-modality Knowledge Conflict in Large Vision-Language Models

   Tinghui Zhu,  Qin Liu, Fei Wang, Zhengzhong Tu, and Muhao Chen

   In arXiv preprint 2024
2. Allerton

   Mitigating Backdoor Threats to Large Language Models: Advancement and Challenges

   Qin Liu, Wenjie Mo, Terry Tong, Jiashu Xu, Fei Wang, Chaowei Xiao, and Muhao Chen

   In Allerton Conference 2024
3. EMNLP

   LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing

   Jiangshu Du, Yibo Wang, Wenting Zhao, Zhongfen Deng, Shuaiqi Liu, Renze Lou, Henry Peng Zou, Pranav Narayanan Venkit, Nan Zhang, Mukund Srinath, and others

   In Proceedings of EMNLP 2024
4. EMNLP

   Securing Multi-turn Conversational Language Models Against Distributed Backdoor Triggers

   Terry Tong, Jiashu Xu,  Qin Liu, and Muhao Chen

   In Findings of EMNLP 2024
5. EMNLP

   Monotonic Paraphrasing Improves Generalization of Language Model Prompting

   Qin Liu, Fei Wang, Nan Xu, Tianyi Yan, Tao Meng, and Muhao Chen

   In Findings of EMNLP 2024
6. NAACL

   Two Heads are Better than One: Nested PoE for Robust Defense Against Multi-Backdoors

   Victoria Graf,  Qin Liu, and Muhao Chen

   In Proceedings of NAACL 2024

   HTML
7. NAACL

   From Shortcuts to Triggers: Backdoor Defense with Denoised PoE

   Qin Liu, Fei Wang, Chaowei Xiao, and Muhao Chen

   In Proceedings of NAACL 2024

   Abstract Code HTML

   Language models are often at risk of diverse backdoor attacks, especially data poisoning. Thus, it is important to investigate defense solutions for addressing them. Existing backdoor defense methods mainly focus on backdoor attacks with explicit triggers, leaving a universal defense against various backdoor attacks with diverse triggers largely unexplored. In this paper, we propose an end-to-end ensemble-based backdoor defense framework, DPoE (Denoised Product-of-Experts), which is inspired by the shortcut nature of backdoor attacks, to defend various backdoor attacks. DPoE consists of two models: a shallow model that captures the backdoor shortcuts and a main model that is prevented from learning the shortcuts. To address the label flip caused by backdoor attackers, DPoE incorporates a denoising design. Experiments on three NLP tasks show that DPoE significantly improves the defense performance against various types of backdoor triggers including word-level, sentence-level, and syntactic triggers. Furthermore, DPoE is also effective under a more challenging but practical setting that mixes multiple types of triggers.

2023

2023

1. preprint

   The Rise and Potential of Large Language Model Based Agents: A Survey

   Zhiheng Xi, Wenxiang Chen, Xin Guo, and others

   In arXiv preprint 2023
2. preprint

   Secrets of RLHF in Large Language Models Part I: PPO

   Rui Zheng, Shihan Dou, Songyang Gao, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin,  Qin Liu, Limao Xiong, Lu Chen, and others

   In arXiv preprint 2023
3. ACL

   Detecting Adversarial Samples through Sharpness of Loss Landscape

   Rui Zheng, Shihan Dou, Yuhao Zhou,  Qin Liu, Tao Gui, Qi Zhang, Zhongyu Wei, Xuan-Jing Huang, and Menghan Zhang

   In Findings of ACL 2023

   HTML
4. ACL

   Characterizing the Impacts of Instances on Robustness

   Rui Zheng, Zhiheng Xi,  Qin Liu, Wenbin Lai, Tao Gui, Qi Zhang, Xuan-Jing Huang, Jin Ma, Ying Shan, and Weifeng Ge

   In Findings of ACL 2023

   HTML

2022

2022

1. ACL

   Flooding-X: Improving BERT’s resistance to adversarial attacks via loss-restricted fine-tuning

   Qin Liu*, Rui Zheng*, Bao Rong, Jingyi Liu, Zhihua Liu, Zhanzhan Cheng, Liang Qiao, Tao Gui, Qi Zhang, and Xuan-Jing Huang

   In Proceedings of ACL 2022

   Abstract Code HTML

   Adversarial robustness has attracted much attention recently, and the mainstream solution is adversarial training. However, the tradition of generating adversarial perturbations for each input embedding (in the settings of NLP) scales up the training computational complexity by the number of gradient steps it takes to obtain the adversarial samples. To address this problem, we leverage Flooding method which primarily aims at better generalization and we find promising in defending adversarial attacks. We further propose an effective criterion to bring hyper-parameter-dependent flooding into effect with a narrowed-down search space by measuring how the gradient steps taken within one epoch affect the loss of each batch. Our approach requires zero adversarial sample for training, and its time consumption is equivalent to fine-tuning, which can be 2-15 times faster than standard adversarial training. We experimentally show that our method improves BERT’s resistance to textual adversarial attacks by a large margin, and achieves state-of-the-art robust accuracy on various text classification and GLUE tasks.

2. COLING

   PlugAT: A Plug and Play Module to Defend against Textual Adversarial Attack

   Rui Zheng, Rong Bao,  Qin Liu, Tao Gui, Qi Zhang, Xuanjing Huang, Rui Xie, and Wei Wu

   In Proceedings of COLING 2022

   HTML

2021

2021

1. ACL Demo

   Textflint: Unified multilingual robustness evaluation toolkit for natural language processing

   Xiao Wang*,  Qin Liu*, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, and others

   In Proceedings of ACL: System Demonstrations 2021

   Abstract Code HTML

   TextFlint is a multilingual robustness evaluation toolkit for NLP tasks that incorporates universal text transformation, task-specific transformation, adversarial attack, subpopulation, and their combinations to provide comprehensive robustness analyses. This enables practitioners to automatically evaluate their models from various aspects or to customize their evaluations as desired with just a few lines of code. TextFlint also generates complete analytical reports as well as targeted augmented data to address the shortcomings of the model in terms of its robustness. To guarantee acceptability, all the text transformations are linguistically based and all the transformed data selected (up to 100,000 texts) scored highly under human evaluation. To validate the utility, we performed large-scale empirical evaluations (over 67,000) on state-of-the-art deep learning models, classic supervised methods, and real-world systems. The toolkit is already available at https://github.com/textflint with all the evaluation results demonstrated at textflint.io.

2020

2020

1. NLPCC

   Learning to generate representations for novel words: Mimic the OOV situation in training

   Xiaoyu Xing, Minlong Peng, Qi Zhang,  Qin Liu, and Xuanjing Huang

   In NLPCC 2020