Ming Li

I am a third-year Ph.D. student in Computer Science at the University of Maryland, advised by Prof. Tianyi Zhou. My research develops post-training, interpretability, and evaluation methods for understanding and improving the behavior of large language models and agentic systems. Over the past few years, I have led multiple projects on data-centric post-training, reasoning behavior analysis, and agentic evaluation, resulting in over 10 first-author papers at top-tier venues including ACL, NeurIPS, ICLR, EMNLP, and NAACL.

Beyond research, I serve as an ACL ARR Area Chair and reviewer for major conferences, and I was honored to receive the 2026 Apple Scholar in AI/ML Fellowship. I have also collaborated with Adobe, Amazon, and Microsoft Research on multimodal learning, LLM reasoning, and post-training.

Research interests

My research focuses on large language models, especially reliable post-training, reasoning interpretability, and agentic systems. Across these areas, I study how to train, understand, and evaluate LLMs so that they can reason more reliably, follow instructions more effectively, and interact with environments more robustly.
(1) Data-centric AI and Post-training: I study data-centric approaches, such as data selection and synthesis, together with algorithmic approaches for improving model adaptation, instruction following, and reasoning (IFD, Superfiltering, Mosaic-IT, Selective Reflection-Tuning);
(2) Model Behavior and Interpretability: I analyze how LLMs and agents reason, follow instructions, overthink, collaborate, and interact with environments, and develop interpretability methods to understand how training and inference shape these behaviors (Layer_Gradient, Gradient_Unified, MiP-Overthinking, ThinkARM, Schoenfeld Reasoning);
(3) Agentic Systems and Evaluation Harnesses: I study how to train more capable agents, how multi-agent systems develop effective coordination behaviors, and how to build evaluation harnesses for agent-environment interaction (ATLaS, Socialization, Superminds Test, ClawEnvKit).

I also study Human-AI Cognitive Alignment: how closely LLM behaviors align with human cognitive patterns, especially in mathematical reasoning and difficulty perception (Difficulty Alignment, Item Difficulty Modeling).

I am always happy to discuss research ideas and collaborations. Feel free to reach out by email.