Working memory capacity and lexical retention in AI-assisted vocabulary learning as predictors of EFL reading...

This study investigates how AI-assisted vocabulary learning translates into English as a Foreign Language (EFL) reading comprehension in Chinese universities-and for whom those gains are strongest. Grounded in the Lexical Quality Hypothesis, retrieval and spacing principles, and capacity-limited accounts of comprehension, we theorize a moderated-mediation in which AI-assisted vocabulary learning engagement and perceived usefulness promote lexical retention, which in turn predicts EFL reading comprehension; working memory capacity moderates the retention-reading link. Cross-sectional data were collected in person from undergraduates in four non-Tier-1 cities (Nanjing, Wuhan, Xi'an, Hangzhou). Measurement combined formative app-log indicators for engagement, a formative lexical-retention composite (breadth, depth, delayed recall), and reflective constructs for working memory capacity and reading comprehension. SmartPLS analyses indicated sound reliability, convergent and discriminant validity, acceptable model fit, and negligible common-method and multicollinearity concerns. Structural tests showed that higher engagement related to stronger lexical retention and better reading; lexical retention functioned as the proximal driver of comprehension; perceived usefulness supported retention (β = 0.19, p < .001) and showed a small direct association with reading (β = 0.08, p = .044); and the retention → reading association was stronger at higher WMC (interaction β = 0.09, p = .005), with simple slopes steeper at +1 SD than -1 SD of WMC. By isolating delayed lexical retention as the mechanism and treating working memory capacity as a boundary condition within a single, parsimonious model, the study suggests shifting practice from raw "time on app" to (a) spaced-review adherence targets and (b) reading-task scaffolds for lower-WMC learners that reduce cognitive load while promoting integration and inference.