LLMs predict forward reading, but fail to explain eye regressions

tallinzen · x · 2026-08-20

Research from NYU and UMass finds that while current language models can accurately predict the forward trajectory of human eye movements during reading, they fail to predict regressive saccades (jumping back) when a reader trips up. This reveals a gap in how LLMs model human reading cognition.

Related event: Language Models Predict Reading Progress but Fail to Explain Regressions(2 posts)→

Original post →

More from Research

Research channel →