Model Reproduces 8 Reading Datasets, Unifying Eye Movements and Comprehension
Updated
Updated · Nature.com · Aug 10
Model Reproduces 8 Reading Datasets, Unifying Eye Movements and Comprehension
2 articles · Updated · Nature.com · Aug 10
Summary
A new hierarchical computational model treats reading as a resource-rational process, choosing fixations, skips and regressions to maximize comprehension while minimizing time and effort.
Three nested control levels—word, sentence and text—were implemented as partially observable decision processes trained with reinforcement learning, linking low-level eye movements to higher-level memory and comprehension goals.
Across 8 human reading datasets, the model reproduced established effects including longer fixations on longer words, more skipping of frequent or predictable words, and more regressions under ambiguity or integration difficulty.
In a new English eye-tracking study, 39 participants read under 30-, 60- and 90-second limits; both humans and the model read faster with more skipping and less regression under tighter time pressure, with comprehension falling.
Ablation tests found the full bounded-memory hierarchy matched human behavior best, while unlimited-memory or myopic variants produced superhuman recall or unrealistic reading patterns, suggesting resource limits and long-term planning are essential.