How L1–L2 typological distance moderates the interplay between cognitive resources and LLM predictors in L2 reading

Abstract This study investigates how L1–L2 typological distance moderates the predictive validity of large language model (LLM)-based metrics in L2 reading, alongside the compensatory role of cognitive resources. Eye-tracking data from Chinese learners with L1 Lao, Burmese, Thai and Vietnamese revealed that surprisal predicted processing difficulty across all groups, whereas contextual entropy effects were group-specific and qualified by cognitive differences. Cognitive abilities explained unique variance only in Vietnamese learners, with verbal short‑term storage (VSTS) showing an independent effect. Additionally, VSTS and reading comprehension accuracy showed significant interactions with typological distance, such that higher learner-level memory and comprehension-performance indices were associated with reduced processing burden under greater L1–L2 divergence. These findings demonstrate that LLM metrics’ validity is constrained by typology and modulated by cognitive resources, supporting adaptive L2 processing models and the Cognitive Resource Coordination Model (Robinson [2001, Second Language Research 17(4), pp. 368–392]).