Unveiling the cognitive features of L2 Chinese reading: Machine learning reveals subskill-specific and general eye-movement patterns.

Traditional reading assessments rely heavily on offline summative outcomes, obscuring the implicit cognitive dynamics of online processing. To render these latent cognitive dynamics empirically tractable, this study employs a data-driven machine learning approach within a non-embedded reading paradigm to reveal six cognitive subskills in 193 learners of Chinese as a second language (L2). Eye-movement data were collected during minimally task-constrained reading phase. The results indicate that eye-movement behavior predicts delayed comprehension performance significantly above chance for five of the six targeted subskills. This establishes the statistical detectability of spontaneous anticipatory processing under ecologically valid conditions. Notably, the subskill more dependents on offline mentalization reveal a boundary condition for online eye-tracking prediction. Beyond subskill-specific prediction, the analysis identifies a three-tier hierarchy of general eye-movement features spanning automatized lexical access, visuomotor spatial planning, and online comprehension monitoring, which constitutes a task-invariant cognitive foundation supporting L2 reading comprehension. Collectively, these findings provide proof of concept for the fine-grained assessment of reading subskills using eye-tracking.