- Researchers at Aalto University, together with international partners, have developed the most accurate model yet of how humans read.
- "For the first time we've used AI methods to understand—not just mimic—how people read," says Professor Antti Oulasvirta from Aalto University.
- The new model is guided by resource rationality—the idea that while reading, we constantly decide where to look next to improve our understanding as much as possible within the time available.
- The model's attention-allocation decisions mirrored real human eye-tracking data, enabling tailoring to different reader profiles.
- A separate study by NYU and UMass found that while AI can explain initial word recognition, it fails to account for integration or rereading.
- Researchers plan to evaluate the model for helping individuals with dyslexia and low language proficiency.
- Earlier models learned from large data sets pairing text snippets with eye-tracking data, then mimicked human behavior, but they lacked true understanding of the content and didn't generalize well across languages or contexts.
- The study from NYU and UMass highlights that while AI models can explain initial word recognition, they struggle with integrating words into larger contexts, especially in complex sentences.
- The researchers used eye-tracking technology to analyze 368 adult readers, focusing on how long participants spent reading and rereading each word of carefully designed sentences.
A new AI model developed by researchers at Aalto University accurately captures human reading behavior, paving the way for personalized text and augmented reality applications. The model, which follows psychological mechanisms of attention allocation, could tailor complex texts to different readers and situations.1
"We placed the model in a world with millions of texts. Then, using AI-based reinforcement learning, we trained it to optimize eye movements so that it truly understands what it reads," explains researcher Oulasvirta. The model adapts to reader characteristics, allowing for customized text presentations, such as making convoluted legal writing more comprehensible for various audiences.
In a separate study, researchers from New York University and the University of Massachusetts Amherst found that while both humans and AI rely on next-word predictions during initial reading, LLMs fail to account for rereading and complex sentence integration. “We found that LLMs can explain how long it takes people to recognize words when their eyes move smoothly forward through a text, but they fail to capture the cases where people have difficulty integrating a word into the larger context of a sentence,” says William Timkey, lead author of the study published in *Proceedings of the National Academy of Sciences* (PNAS).
The study analyzed 368 adult readers using eye-tracking technology, revealing that AI models underpredict the difficulty humans face with complex sentences. “The predictability of a word really doesn’t even come close to explaining just how much time we spend on difficult words and garden-path sentences,” Timkey adds. Researchers emphasize the need for AI models that better reflect human cognitive processes to bridge the gap in understanding.89
“The Aalto model, guided by resource rationality, trains via reinforcement learning to optimize eye movements, mirroring human gaze decisions. In contrast, the NYU/UMass study of 368 readers found AI predicts initial word recognition but not the integration step, where 20% of eye movements are backward.”
Neuroscience News
