Rongwu Xu (许融武)

Research

I am an artificial intelligence (AI) researcher with pretty interdisciplinary interests. I try to understand how AI's design and its interactions with humans can lead to unexpected behaviors and increased societal risks. I approach this work through the lens of behavioral experiments, machine learning (ML), interpretability tools, and psychology. I publish my findings in the Natural Language Processing (NLP), AI, and ML communities.

Currently, I am interested in the following research topics:

1. AI Safety and Alignment: Identifying potential safety and ethics risks associated with AI R&D and developing strategies to align AI systems with human values, behaviors, and expectations.
2. Machine Behavior: Investigating the similarities and differences between AI models and human behaviors, and utilizling psychology-inspired experiments to test and understand machines.
3. AI and Psychology: Studying both the understanding the psychological impacts of AI systems on humans (Psychology of AI, a subset of Psychology of Technology) and the application of AI in psychological research (AI for Psychology, a subset of AI for Science).

My other general interests include model evaluation and real-world applications of such models.

News

AI/NLP research can be challenging for newcomers. If you're interested in my work or have ideas to explore, I'd be happy to guide you. We can work on submitting papers to top venues. Feel free to drop me an Email if interested.

Selected Publications

• "Nuclear Deployed!": Analyzing Catastrophic Risks in Decision-making of Autonomous LLM Agents
  Rongwu Xu*, Xiaojian Li*, Shuo Chen*, Wei Xu
  Working Paper
  [Paper][Project Page][Code][X Post]
  
• Knowledge Conflicts for LLMs: A Survey
  Rongwu Xu*, Zehan Qi*, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, Wei Xu
  EMNLP 2024 [CORE A*]
  [Paper][Code][机器之心][Talk (Chinese)][Slides][Poster][X Post]
  
• How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States
  Zhenhong Zhou, Haiyang Yu, Xinghua Zhang, Rongwu Xu, Fei Huang, Yongbin Li
  EMNLP 2024 Findings [CORE A*]
  [Paper][Code][Poster]
  
• The Earth is Flat because...: Investigating LLMs' Belief towards Misinformation via Persuasive Conversation
  Rongwu Xu, Brian S. Lin, Shujian Yang, Tianqi Zhang, Weiyan Shi, Tianwei Zhang, Zhixuan Fang, Wei Xu, Han Qiu
  ACL 2024 Oral [CORE A*]
  🏆 Outstanding Paper Award [Certificate]
  [Paper][Project Page][Code][机器之心][Video][Poster]
  

* Equal Contribution, ^ Advising Role

Awards

• Most Recognized Research Outcomes at Tsinghua University Nomination (清华大学最受师生关注的年度亮点成果提名, Top 2 at IIIS), 2024
• Tsinghua University Excellent Teaching Assistant (清华大学优秀助教, Top 2%), 2024
• National Scholarship (国家奖学金, Top 1%), 2024
• Tsinghua University Outstanding Student Cadre (清华大学优秀学生干部, Top 1.5%, 121 out of 9000+), 2024
• ACL 2024 Outstanding Paper Award (Top 0.79%, 35 out of 4407), 2024
• Tsinghua-Yangtze River Delta International R&D Community Talent Scholarship, 2023
• Tsinghua University Overall Excellence Scholarship (Top 10%), 2023
• Tsinghua University Overall Excellence Scholarship (Top 10%), 2022
• Tsinghua University Technological Innovation Excellence Scholarship, 2020
• Tsinghua-Panasonic Scholarship (Top 10%), 2019
• Outstanding Volunteers in Beijing, 2018

Talks

• “做科研的选择” (在2024年交叉信息院开学典礼上作为在校生代表的发言), IIIS, Tsinghua University, Aug 2024
  [Speech (Chinese)]
• Knowledge Conflicts for LLMs, NICE (w. Soochow University), Jul 2024
  [2-Hour Talk (Chinese)][Slides]
  
• LLMs Safety Vulnerabilities via Contextual Misinformation, IIIS, Tsinghua University, May 2024
  [3-min Clip]

Professional Service

• Peer Reviewer, ACL Rolling Review, 2024-Present
  Tracks: Ethics, Bias, and Fairness, Human-Centered NLP (2025-), NLP Applications, Resources and Evaluation