Brain study reveals prediction vs. feedback in language learning with transformers

A new study from researchers at South China Normal University and others sheds light on how the brain learns new languages by disentangling the roles of predictive and feedback signals. Over seven days, 102 adults learned an artificial language while undergoing fMRI scans. The task involved corrective feedback, mimicking real-world language learning. To interpret the neural data, the team trained transformer-based neural networks (similar to models like GPT) on three objectives: pure prediction (next-word), pure feedback (error correction), and a combined objective. They then compared the internal representations of these models to the brain activity patterns of the learners.

Strikingly, the prediction-focused model accounted for the largest share of shared neural variance across all participants — even though the human task itself was feedback-based. This suggests that prediction forms a common neural architecture that most learners rely on. However, the feedback model was the best predictor of individual differences in how well participants generalized their learning on Day 7. The findings point to a dual-signal mechanism: prediction drives group-level language processing, while feedback shapes personalized learning trajectories. This work has implications for AI-based language tutoring and understanding neural basis of learning.