Russia Patents AI That Reads Texts Like a Human

Researchers have been granted patents for a system that predicts human gaze during reading. It forecasts how a person will read text in English, Korean, German, and ten other languages, based on a single neural network architecture combining a multilingual model and a transformer trained on eye movement recordings from native speakers. The system generates synthetic gaze fixation sequences and evaluates them by word skip probability, number of fixations during first reading, and other criteria. Ilya Pershin, head of the AI in Medicine Lab at Innopolis University, explained that synthetic data is often the only solution when real eye-tracking recordings are scarce, making training cheaper and enabling previously impossible tasks. In one study, this approach improved gaze prediction accuracy on X-ray images by 20–30%. The second patent covers a unified protocol for evaluating eye movement generation during reading, comparing gaze trajectories by spatial and temporal characteristics, and analyzing features at whole-text, word, and sentence levels. The third patent describes a method for training language models with synthetic gaze trajectories, integrating visual attention data into reinforcement learning from human feedback. Ivan Stebakov, a lab specialist, said that a reward model is trained on real human preferences to automatically evaluate responses from a large language model, accelerating training by 1.5–2 times, reducing computational costs, achieving 93% accuracy in reproducing human attention patterns, and working even for low-resource languages.