Local Predictability Changes Modulate Reading in Naturalistic Texts

Reading unfolds incrementally, and word predictability varies from moment to moment across a text. Yet most accounts treat predictability as a static property of individual words, overlooking how recent changes in predictability may shape ongoing comprehension. The present study addresses this gap by investigating how fluctuations in predictability shape reading behavior and how these effects vary with language proficiency. Using a pretrained GPT-2 model to derive word-by-word predictability estimates from naturalistic texts, Chinese–English bilinguals read English newspaper excerpts in a self-paced reading task. The results showed that readers with smaller vocabularies exhibit greater surprisal effects. Critically, for these readers, the magnitude of the surprisal effect at Wordn was modulated by recent predictability history: effects were greater following greater increases in predictability from Wordn-2 to Wordn-1. This pattern suggests that predictive engagement is adaptively adjusted in response to recent contextual change, consistent with predictive coding accounts but inconsistent with volatility-tracking accounts. Overall, these findings highlight the dynamic nature of predictive processing during reading and suggest how individual differences in lexical knowledge modulate this adaptation.