Human Attention for Large Language Model Training

Human readers do not distribute attention evenly. Some words invite long fixations; others pass almost unnoticed. This thesis asks whether those patterns can become a useful signal for language models without mistaking familiar lexical effects for cognition.

The work builds controlled analyses that separate gaze measurements from confounders including word length and frequency. It then tests gaze-informed training and reward-model interventions, alongside mechanistic diagnostics of attention heads.

The project was completed in collaboration with Telefónica I+D and the Universitat Politècnica de Catalunya. It received a grade of 10/10 and an honours distinction.

The full thesis will be linked here when a public copy is available.