Toward A Foundation of Reprogramming Large Language Models by Chi Han

Abstract: Large language models (LLMs) have achieved strong performance in domains covered by pre-training corpora. However, once pre-trained, their internal computations are largely treated as 
fixed black-boxes, and systematically modifying them remains an open challenge. As a result, 
when LLMs exhibit limitations in open-world settings, such as extreme-scale contexts, novel 
scenario generation, and scientific discovery, we lack principled methods to intervene at the 
internal mechanistic level where these failures originate. In this talk, I present a quest toward a 
principled foundation for LLM reprogramming by uncovering governing principles of their 
internal computations and developing methods for post-hoc reprogramming of LLMs.
My approach integrates two complementary components: theoretical analyses of how LLMs 
process context and generate decisions, and empirical methods that reprogram internal 
architectures with minimal additional resources. First, I present LM-Infinite, which provides a 
mechanistic explanation of context-length failure and enables generalization to contexts 
exceeding 200 million tokens without retraining. Second, in LM-Steer, I show that word 
embeddings act as controllable steers for generation, enabling efficient, interpretable, and 
transferable control over LLM outputs. Third, extending beyond language, I introduce a 
modular chemical language model that incorporates domain-grounded representations to 
support synthesis-aware molecular reasoning and drug discovery. Together, my research sheds 
light on principled mechanisms for making LLMs more interpretable, adaptable, and reliable 
across scientific disciplines and real-world settings.

Bio: Chi Han is a final-year Ph.D. candidate in Computer Science at the University of Illinois 
Urbana-Champaign (UIUC), where he works in the NLP group under the supervision of Prof. 
Heng Ji. He received his undergraduate degree from Tsinghua University, China, through the 
Yao Class program. His research has led to first-author publications at top venues including 
NeurIPS, ICLR, TMLR, ACL, and NAACL, which received Outstanding Paper Awards at NAACL 
2024 and ACL 2024, as well as the Best Demo Award (1st place) at the NSF Summit for AI 
Institutes Leadership (SAIL). His research has been supported by the IBM PhD Fellowship, 
Amazon AICE PhD Fellowship, the Mavis Future Faculty Fellowship, and Capital One seeding 
grant. His research interests focus on developing theoretical foundations and empirical 
methods for post-hoc reprogramming of large language models (LLMs). His work addresses 
intrinsic limitations of LLMs across multiple domains, including context length, decisionmaking transparency, generation control, and scientific discovery, and achieves strong 
performance across downstream tasks