LEXIC injects word-level difficulty signals into gaze-only eye-tracking models

Researchers propose LEXIC, a lightweight extension for gaze-only eye-tracking models that improves reading comprehension prediction by injecting precomputed word-level difficulty signals without using pretrained language models. The approach builds on the EyeBench AhnCNN baseline and introduces two mechanisms—LEXIC-Concat and LEXIC-Res—to incorporate GPT-2 surprisal, word frequency, and word length into per-fixation inputs.

On the OneStop reading comprehension task with K=5 seed-ensemble training across ten folds, both mechanisms yield statistically consistent AUROC gains of +1.8 to +2.2 percentage points on Unseen Text (Wilcoxon p <= 0.065). LEXIC-Concat additionally lifts performance on Unseen Reader by +2.9 percentage points (p = 0.010), whereas LEXIC-Res shows a smaller, non-significant gain of +1.8 percentage points due to calibration issues with out-of-distribution readers.

The study highlights an architectural boundary in gaze-only modeling, demonstrating that lightweight conditioning can bridge the gap between chance-level performance and text-aware models, though transferability to unseen readers remains challenging.