Summary:
- The article explores how large language models (LLMs) can be utilized as predictive proxies for human cognitive processes, specifically in the context of eye-tracking data during sentence reading.
- It highlights research demonstrating that "surprisal"—a measure of how unexpected a word is to a model—correlates significantly with human reading patterns, such as fixation duration and the likelihood of regressive eye movements (rereading).
- The findings suggest that the internal mechanisms of LLMs provide a powerful computational framework for testing theories of human language processing and cognitive load.