Researchers have investigated whether language model representations encode signals related to human reading times. By using regularized linear regression on eye-tracking data across five languages, they compared model layer representations against predictors like surprisal. The study found that early model layers were better at predicting initial reading measures, suggesting that low-level representations capture human-like processing signatures. However, for later reading measures, surprisal remained a stronger predictor, and the optimal predictor varied by language and eye-tracking metric. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON This is a research paper detailing a study on language model representations and human reading times.