Ming Li

I am a Ph.D. candidate in Computer Science at the University of Maryland, advised by Prof. Tianyi Zhou. I study how to train LLMs, understand their reasoning and behavior, and rigorously evaluate the capabilities and reliability of models and agentic systems.

My research experience includes internships at Amazon, Microsoft Research, and Adobe, and a visiting research appointment at MBZUAI. I also serve as an ACL ARR Area Chair and reviewer for major conferences.

2026 Apple Scholar in AI/ML · 2× ACL Oral · ACL’26 SAC Highlight

I expect to graduate in Spring 2027 and am seeking Research Scientist / Applied Scientist / Research Engineer roles in industry. Download my CV or get in touch.

Research

(1) Data-Centric LLM Post-Training: I develop methods to select and reuse high-value training signals, from instruction-data selection (IFD, Superfiltering) to critical-step agent tuning (ATLaS). I also use layer-wise gradient analysis to understand how instruction and reasoning data shape post-training and what makes training data effective.

(2) Reasoning & Model Behavior: I study how training and inference shape reasoning structure and reliability, using functional analysis of reasoning traces (ThinkARM, ACL’26 Oral & SAC Highlight) and layer-wise gradient analysis (Fast vs. Slow Thinking, ACL’25 Oral). I also investigate reasoning failures such as overthinking under missing premises and what RL improves in visual reasoning.

(3) AI Evaluation & Reliability: I design controlled evaluations to probe capability boundaries and unreliable judgments, including collective behavior in AI agent societies (Moltbook; Science News interview) and rhetorical sensitivity in AI reviewers. Related directions include multimodal capabilities (ColorBench), efficient evaluation (BenTo), and agent evaluation environments (ClawEnvKit).

I also explore Human-AI Cognitive Alignment, especially how model reasoning and difficulty judgments relate to human cognition: Difficulty Alignment, Item Difficulty Modeling, and Episode2Difficulty.

I am always happy to discuss research ideas and collaborations.