Cognitive Processes of Probabilistic Prediction in Reading: Language Model Surprisal Across Model Sizes, Token Granularity, and Reading Paradigms

Abstract

Surprisal, the negative log probability of a word given its context, is the dominant computational metric for quantifying reading difficulty and a common item difficulty estimator in reading research. Yet how the language model family, surprisal granularity, and corpus type jointly shape the surprisal–reading time link remains unclear. We conducted a secondary analysis of two public English reading corpora: the Natural Stories Corpus (self-paced reading, 181 readers) and the Provo Corpus (eye tracking with cloze norms). Surprisal was computed from GPT 2 (Small, XL), Llama 2 (7B, 13B), and Llama 3 (8B), together with a 5-gram baseline and human cloze norms. Linear mixed-effects models tested the baseline contribution of neural surprisal, inverse scaling within model families, word-level versus sub-word aggregation, and the linking function shape. Neural surprisal contributed reliable variance above a strong baseline in both corpora. A clear inverse scaling pattern emerged: GPT 2 Small produced the largest fit improvements, exceeding GPT 2 XL, Llama 2 13B, and Llama 3 8B. Word-level aggregation outperformed sub-word aggregation, especially for measures of early lexical access. Non-parametric analyses supported an approximately linear linking function, and cloze norms carried information not fully captured by neural surprisal. These findings show that larger models are not automatically better cognitive models of reading and that surprisal granularity is not a neutral analytic choice.