Jieyu Zhao · 赵洁玉

🐾 Research Papers @ LIME Lab

My research is on current AI models and agents: how quickly they adapt, what they can really do, how they act in the world, what emerges as systems get more capable, and how we evolve alongside them.

• Adaptation. When does adaptation actually stick, and when does it slip? RFT · Alignment · In-context · Curriculum Efficient post-training (RFT, RLHF) and in-context methods that let LLMs and agents pick up new skills and domains fast. Lately I'm especially curious about the dynamics of adaptation itself: when it sticks, when it slips, and what role world models might play in making it reliable. See: AdaRFT [arXiv 25] , ERL [ICLR 26 Workshop] , WildFeedback [ACL 26] , GRAVITY [EACL 26]
• Agents. How do agents work in messy environments, and how do they help domain experts? CUA · Tool use · Reward modeling · Domain agents I work on agents on two fronts. First, the foundations: computer-use agents (CUA), VLA models, and tool-using LLMs, with a focus on what breaks and what closes the loop: action spaces, reward signals, error recovery, and deployment safety. Second, building agents that help users from other domains, especially medicine and drug discovery, where domain knowledge and trust matter as much as raw capability. See: CoAct-1 [ICLR 26] , ExeVRM [arXiv 26] , OS-Blind [arXiv 26] , MedCopilot [ACL 26 Demo] , DrugAgent [AI4Research @ AAAI 25]
• Emergence. What emerges as systems get more capable? Hallucination · Reward hacking · Data attribution · Multi-agent Emergence cuts both ways. I'm interested in side-effects of capable models (e.g., hallucinations, refusal degradation, reward hacking). And I'm drawn to the unknowns of multi-agent interaction: when agents cooperate, conflict, mediate, or negotiate, what new behaviors appear? Useful capabilities as much as risks. See: Hallucination Tax [EMNLP 25 Findings] , Unsafe Data Attribution [ICML 26] , ProMediate [ACL 26 Findings]
• Evaluation. How do we know what models can really do? Probing · Multimodal · Cultural · Trustworthiness Going beyond benchmark numbers to probe what models really know (reasoning, multimodal grounding, cultural intelligence, agent competence, trustworthiness) and where they quietly fail, including the biased and unsafe failures most likely to cause real-world harm. See: Probing VLMs [EACL 26] , CQ-Bench [arXiv 25] , SEA [COLM 25] , CulturalPersonas [EMNLP 25 Findings] , TrustGen [ICLR 26]
• Co-evolution. How does AI change humans, not just the other way? Human-AI interaction · Sociotechnical · Trust · Long-term influence Most alignment work studies how human feedback shapes AI (e.g., RLHF); I'm equally interested in the reverse: how living with capable AI reshapes human cognition, creativity, work habits, and social norms, and how to make that co-evolution visible, so it isn't something that just happens to us. More to come. This direction is in early exploration.

• [May 2026] 🎤 Gave a talk at ServiceNow Research.
• [May 2026] 📢 Call for submissions to the Secure and Trustworthy LLM Workshop!
• [Apr 2026] 🎤 Gave a tea talk at Hippocratic AI.
• [Apr 2026] 🎤 Gave a tea talk at Mila.
• [Mar 2026] 💻 The code for ERL is available here — play with it!
• [Mar 2026] 📑 New paper out: Video-Based Reward Modeling for Computer-Use Agents. Use it to evaluate the performance of your agents!
• [Mar 2026] We're so happy to welcome Jiaqi Deng to our group!
• [Mar 2026] 📑 Two papers at EACL: What's Missing in Vision-Language Models? Probing Their Struggles with Causal Order Reasoning and GRAVITY: A Framework for Personalized Text Generation via Profile-Grounded Synthetic Preferences. Drop by if you're attending!
• [Mar 2026] 🍋‍🟩 Reopened the open office hour!
• [Feb 2026] 🎉 ICLR 2026 accepted: CoAct-1: Computer-using Multi-agent System with Coding Actions, Doxing via the Lens: Revealing Location-related Privacy Leakage on Multi-modal Large Reasoning Models, and On the Trustworthiness of Generative Foundation Models. Congrats to all collaborators!
• [Feb 2026] 📑 New paper: Experiential Reinforcement Learning. Please upvote ⬆️ if you like it!
• [Feb 2026] 🎓 Congrats to Ziyi for passing her PhD qualifying exam!

• Talk · CPIL, McGill University · 2026-05
• Invited Talk · ServiceNow · 2026-05
• Tea Talk · Mila · 2026-04
• Invited Talk · Hippocratic AI · 2026-04
• Invited Talk · Microsoft Montreal · 2026-03
• IVADO Workshop on Cognitive Basis of Reasoning (in Minds and AI) · IVADO · 2026-01
• IVADO Agent Workshop · IVADO · 2025-11
• Building Personalized AI Assistants: From Task Execution to Human Alignment · UPenn NLP Seminar (CLunch) · 2025-10
• Guest Lecture — CSCI 697 · USC · 2025-09
• UN Women Academic Conference (AI and Gender) · United Nations Entity for Gender Equality and the Empowerment of Women · 2025-08

🏆 Awards (selected)

• SAC Highlight Paper Award , EMNLP 2025 , 2025
• 100 Brilliant Women in AI Ethics , 2025
• Capital One Research Award , 2024
• EMNLP Outstanding AC Award , 2024
• USC WiSE Gabilan Assistant Professorship , 2023
• NSF CI Fellowship , 2021
• MIT EECS Rising Star , 2021
• Microsoft PhD Fellowship , 2020
• SoCalNLP Symposium Best Poster Award , 2018
• UCLA Graduate Division Fellowship , 2017
• EMNLP Best Long Paper Award , 2017
• Outstanding Graduate of Beijing City , 2016

🎓 Education & Experience

• Gabilan Assistant Professor, Thomas Lord Department of Computer Science, USC · 2023.8 – present
• Microsoft Productivity Scholar, Microsoft Research · 2025.2 – 2025.10
• NSF Computing Innovation Fellow, University of Maryland, College Park · Advised by Hal Daumé III · 2022.1 – 2023.6
• Ph.D. in Computer Science, University of California, Los Angeles · Advised by Kai-Wei Chang · 2017.9 – 2021.12
• Research Intern, Google, NYC · 2021.6 – 2021.9
• Research Intern, Allen Institute for AI (AI2), Seattle · 2020.9 – 2020.12
• Research Intern, Microsoft Research, Redmond · 2020.6 – 2020.9
• Research Intern, Microsoft Research, Redmond · 2019.6 – 2019.9

✉️ Contact