Computational Modeling of Eye Movements in Reading Across Multiple Regimes – Theory and Practice

Abstract: Eye movements in reading are widely considered as a window into the reader’s cognitive processes. In this talk, I will explore non-ordinary reading as a frontier for cognitive modeling and psycholinguistics, emphasizing its potential to deepen our understanding of language processing and reader behavior in humans, and study the alignment between text processing in humans and LLMs. In the first part of the lecture, I will present an in-depth analysis of how human eye movement patterns differ between first and repeated readings, leveraging a large-scale eye movement in reading dataset. This analysis extends and challenges prior work, offering a detailed empirical picture of eye movements during repeated reading and their underlying cognitive processes. The second part will focus on empirical investigations of surprisal theory in goal-driven reading and repeated reading. I will discuss our findings that reveal robust misalignments between LLMs and human text processing that challenge current theories of human text processing. Our results raise critical questions about the ability of such models to estimate cognitively relevant quantities and provide meaningful insights into human linguistic behavior. Finally, I will discuss the theoretical and practical challenges these findings present, addressing their implications for both cognitive modeling and the broader application of language models in psycholinguistic research.