Language · Intelligence · Modeling · Evaluation LIME Lab is part of the USC NLP Group, led by Prof. Jieyu Zhao. We work on two sides of AI models and agents — modeling them to be more capable, aligned, and reliable in the real world, and designing evaluations that surface where they fall short, with a strong human-centered thread throughout.
ProMediate accepted to ACL 2026 Findings.
Data attribution paper for filtering unsafe training data accepted to ICML 2026.
ICLR 2026 accepted: CoAct-1, Doxing via the Lens, and TrustGen.
ERL code is now available — play with it.
New paper out: Video-Based Reward Modeling for Computer-Use Agents.
Welcomed Jiaqi Deng to the group.
Two papers at EACL: "What's Missing in Vision-Language Models?" (Causal Order Reasoning) and GRAVITY.
Reopened the open office hour. 🍋🟩
New paper out: Experiential Reinforcement Learning.
Ziyi passed her PhD qualifying exam.
Treat past episodes as a retrievable experience bank rather than gradient updates only — letting the policy reason about what worked last time before deciding what to try next.
Taiwei Shi, Sihao Chen, Bowen Jiang, Linxin Song, Longqi Yang, Jieyu Zhao
Find which training examples are pushing a model toward unsafe behavior, by attributing model outputs back through denoised representations — and filter them before they do damage.
Yijun Pan, Taiwei Shi, Jieyu Zhao, Jiaqi W. Ma
Most agent benchmarks are single-turn task completion. ProMediate puts agents in the middle of multi-party negotiations and asks whether they can read the room, mediate, and act proactively.
Ziyi Liu, Bahar Sarrafzadeh, Pei Zhou, Longqi Yang, Jieyu Zhao, Ashish Sharma
Align LLMs with the messy, in-the-moment feedback users actually leave during real interactions — not the curated preference pairs that look clean on paper.
Taiwei Shi, Zhuoer Wang, Longqi Yang, Ying-Chun Lin, Zexue He, Mengting Wan, Pei Zhou, Sujay Jauhar, Sihao Chen, Shan Xia, Hongfei Zhang, Jieyu Zhao, Xiaofeng Xu, Xia Song, Jennifer Neville
A computer-using agent that doesn't just point and click — it writes code as a first-class action, finishing tasks faster and with higher success than GUI-only agents.
Linxin Song, Yutong Dai, Viraj Prabhu, Jieyu Zhang, Taiwei Shi, Li Li, Junnan Li, Silvio Savarese, Zeyuan Chen, Jieyu Zhao, Ran Xu, Caiming Xiong
We study AI systems beyond static benchmarks: how they adapt, act, fail, and become trustworthy in real-world settings.
We study how models adapt through post-training, reinforcement finetuning, user feedback, personalization, and curriculum design.
ExploreWe build and evaluate agents that use computers, tools, and domain knowledge to plan, act, recover, and collaborate.
ExploreWe design evaluations that reveal what models actually know, understand, and fail to do across reasoning, multimodal, cultural, and agentic tasks.
ExploreWe audit failures that matter in real use: bias, hallucination, privacy leakage, unsafe data influence, and brittle model judgments.
Explore100 Brilliant Women in AI Ethics
EMNLP SAC Highlight Award Huanxin Sheng
Analyzing Uncertainty of LLM-as-a-Judge: Interval Evaluations with Conformal Prediction
Capital One Student Fellowship Ziyi Liu
Capital One Student Fellowship Priyanka Dey
EMNLP Outstanding AC Award
USC–Capital One Research Award
WiSE and EMNLP Travel Grants Ziyi Liu
Amazon Responsible AI Funding
Gabilan Assistant Professor Fellowship
SoCal NLP Symposium Best Poster Award
Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods
EMNLP 2017 Best Long Paper Award
Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints
