Predicting L2 Eye Movements with Generalization and Reader-Specific Adaptation

ACL ARR 2026 January Submission186 Authors

Keywords: second language reading, eye movements, token-level prediction

Abstract: We cast L2 reading as a predictive generalization problem: can word-by-word eye movements be predicted for unseen readers and unseen texts? Using a large naturalistic eye-tracking corpus, we model nine eye-movement measures as token-level targets and evaluate cross-domain generalization under three settings: unseen texts, unseen readers, and both. To test whether learner variables provide transferable information beyond identity shortcuts, we compare true learner profiles against a column-wise permutation control. We further study the sample efficiency of personalization via a two-stage k-shot residual calibration method that adapts a general predictor to new readers. Our framework establishes a reproducible benchmark for L2 reading prediction and links cognitive variables to out-of-sample performance.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Machine Learning for NLP, Interpretability and Analysis of Models for NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: Chinese (L2), Vietnamese (L1), Thai (L1), Lao (L1), Burmese (L1)

Submission Number: 186