LEXIC boosts gaze-only eye-tracking with lightweight complexity injection

For the first time, a gaze-only model can meaningfully predict reading comprehension from eye movements, thanks to LEXIC (Lightweight Eye-tracking eXtension via Injected Complexity). The research team—spanning Korea and the US—found that existing gaze-only models performed at chance on the EyeBench benchmark, while text-aware models using pretrained language models reached 56–63% AUROC. LEXIC bridges this gap by injecting three precomputed word-level difficulty signals (GPT-2 surprisal, word frequency, and word length) into the per-fixation encoding of a lightweight CNN baseline called AhnCNN. Two variants were tested: direct concatenation (LEXIC-Concat) and a residual mechanism (LEXIC-Res) where a small head predicts typical-reader gaze response and conditions the encoder on deviation.

On the OneStop reading comprehension task, using K=5 seed-ensemble training across ten folds, LEXIC-Concat achieved statistically significant (+2.9 percentage points, p=0.010) gains on Unseen Reader evaluation, and both variants showed consistent +1.8 to +2.2 percentage point improvements on Unseen Text (Wilcoxon p ≤ 0.065). The residual variant struggled with out-of-distribution readers due to calibration to training readers. Overall, LEXIC demonstrates that lightweight, language-model-free conditioning can substantially improve gaze-only models, making reading comprehension prediction accessible without heavy NLP infrastructure. The paper has been accepted to APCCAS 2026.