Aalto University Researchers Build AI Reading Model for AR and Text Personalization

Researchers at Aalto University (Finland), together with international partners, have built the most accurate computational model of human reading to date. The new system uses reinforcement learning, a type of AI from the world of robotics, to reproduce and explain the decisions a reader makes while moving through a text.

Why is human-style reading so hard to model?

Eye movements during reading are among the hardest cognitive processes to model: a person makes dozens of unconscious decisions every second. The researchers reproduced three key types of these movements: saccades (rapid jumps of the gaze between words), fixations (pauses to process information), and regressions (returns to earlier text when a passage is not understood).

Previous models relied on statistics and fixed psycholinguistic rules, and handled individual differences between readers poorly. The RL agent works differently: it learns through trial and feedback, gradually building an adaptive strategy that is sensitive to context, including word length and frequency, syntactic complexity, and the reader's goal.

• Developer: Aalto University (Finland) together with international partners
• Method: reinforcement learning
• Subject: eye saccades, fixations, and regressions during reading
• Result: the most accurate computational model of reading as of August 2026
• Applications: text personalization, augmented reality, accessibility, and diagnostics

"The new model uses reinforcement learning, a type of AI from the world of robotics, to explain and reproduce the decisions readers make as they move through a text," according to

Aalto University's description of the study.

What will the new model be used for?

Content personalization: if a system knows how a specific reader processes text, it can dynamically adjust sentence length, formatting, and emphasis. For educational platforms, this affects how well material is absorbed; for media, it affects reader depth and retention.

Augmented reality: AR glasses and headsets need to know where the user is looking right now so they can overlay information precisely at the point of focus, without delay or offset. Hardware eye trackers are accurate but expensive and power-hungry. A software-level predictive model reduces the dependence on specialized hardware.

Accessibility and diagnostics: in people with dyslexia or attention deficits, reading strategies differ systematically from the norm. An accurate model can identify these differences and help design text interfaces that are easier for such users to read.

According to Aalto University's description of the study, applying RL made it possible for the first time to reproduce the decision-making mechanism behind reading itself, rather than an averaged pattern of eye movements drawn from a dataset.

What this means

Academic models of reading have existed since the classic psycholinguistic research of the 1980s. The shift to an RL approach changes their nature: from descriptive, they become explanatory and potentially adaptive to the individual. For developers of AR interfaces, EdTech platforms, and publishers, this is a new personalization tool that does not require costly sensors. The open question is how well the model generalizes across languages, writing systems, and text types.

Frequently asked questions

What is reinforcement learning in the context of reading?

Reinforcement learning is an AI method in which an agent learns by trial and error, receiving feedback from its environment. The same technology is used in robotics and in training game-playing agents; Aalto University researchers applied it to model how a reader makes decisions while moving through a text.

How does the new model differ from previous ones?

Previous models produced averaged predictions based on statistics and fixed rules. The new one reproduces an adaptive decision-making mechanism that is context-dependent and closer to the behavior of an individual person than to an averaged group of test subjects.