Assessing Cognitive Effort in L2 Idiomatic Processing: An Eye-Tracking Dataset

The dataset's real contribution isn't just measuring cognitive load in idiom processing, but validating that cheap hardware (60 Hz Tobii) can reliably capture the fine-grained reading events that linguists and ML researchers need to ground language model evaluation in actual human processing patterns rather than task accuracy alone.

This connects directly to the interpretability work from early May, particularly the encoding probe paper that asked what language models actually encode versus what we assume they do. Where that work reconstructed model internals from linguistic features, this dataset provides the human ground truth for what 'correct' processing of figurative language looks like cognitively. It also echoes the memory and procedural execution papers: models fail not just on reasoning but on tracking the cognitive steps humans take. Having empirical eye-tracking data on how proficiency levels affect idiom interpretation gives researchers a measurable target for what training objectives should optimize toward.

If papers citing this dataset appear by Q4 2026 showing that language models trained on eye-tracking-derived loss functions outperform standard supervised baselines on idiom comprehension benchmarks, that confirms the data has real training value. If adoption stays limited to academic psycholinguistics labs without NLP uptake, the bridge between the two fields remains theoretical.