Tsinghua Yao Alumnus Han Chi Joins UW-Madison; ACL/NAACL Outstanding Paper Awards
Tsinghua Yao Class alumnus and UIUC PhD Han Chi will join UW-Madison as assistant professor in 2027, known for award-winning research on LLM internal mechanisms including LM-Infinite for infinite context length and LM-Steer for controllable generation, plus AI for Science work on modular chemical language models.
Background
Dr. Han Chi (韩迟), a graduate of Tsinghua University's Yao Class, completed his Ph.D. in Computer Science at the University of Illinois Urbana-Champaign (UIUC) under Professor Heng Ji. He recently announced on X that he will join the Department of Computer Science at the University of Wisconsin–Madison (UW–Madison) as an assistant professor starting Fall 2027.
Research Focus
Han Chi's research centers on three interconnected areas: understanding the internal mechanisms of large language models (LLMs), developing methods for controllable model behavior, and extending LLM capabilities across diverse tasks and scientific domains. Rather than simply scaling data and compute, he seeks to uncover how models form knowledge, process information, and can be guided to surpass their inherent capability boundaries.
LM-Infinite: Extending Context Length Without Retraining
Published at NAACL 2024 and honored with the Outstanding Paper Award, LM-Infinite addresses the problem that LLMs trained on limited context lengths (e.g., 2K or 4K tokens) degrade when faced with longer inputs. The method requires no retraining of model parameters. Experiments show that models trained on 2K or 4K token contexts can generalize to inputs of up to 200 million tokens while maintaining perplexity performance. This opens new possibilities for processing scientific literature, large codebases, and long dialogues.
LM-Steer: Controllable Generation via Linear Transformation
Also an Outstanding Paper Award winner at ACL 2024, LM-Steer explores controllable text generation. The team discovered that applying a linear transformation to the model's output embedding vectors can steer generation style — such as sentiment — while only learning a small number of extra parameters, leaving the original model largely unchanged. This lightweight approach enables fine-grained control without large-scale model modification.
mCLM: Modular Chemical Language Model for AI-Driven Molecular Design
Extending into AI for Science, Han Chi served as co-first author on mCLM, a modular chemical language model accepted as an Oral presentation at ICLR 2026 and awarded the NSF SAIL 2025 Best Demo Award (First Place). mCLM decomposes molecules into functional chemical modules, allowing the model to design target molecules while simultaneously considering synthesizability via automated synthesis techniques. This work aims to bridge the gap between AI-generated molecular designs and real-world manufacturability.
Awards and Recognition
IBM PhD Fellowship
Amazon AICE PhD Fellowship
Mavis Future Faculty Fellowship
Continuous Tsinghua Yao Class Scholarships
NAACL 2024 Outstanding Paper Award (LM-Infinite)
ACL 2024 Outstanding Paper Award (LM-Steer)
ICLR 2026 Oral Presentation (mCLM)
NSF SAIL 2025 Best Demo Award First Place (mCLM)
Future Directions
As he transitions to an independent faculty role at UW–Madison, Han Chi plans to continue investigating LLM internal mechanisms and capability boundaries, pushing toward broader applications across scientific and practical domains.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
