Fig. 1: Experimental procedure of the dataset and the analyses pipeline.
From: Increasing alignment of large language models with language processing in the human brain
a, Experiment procedure. Participants read five English articles sentence-by-sentence while inside the fMRI scanner with concurrent eye-tracking. b, LLMs of different sizes with and without fine-tuning are employed in the study. c, Analysis pipeline. The attention matrices of each layer of the LLMs for each sentence in the experimental stimuli were averaged over attention heads and aligned with eye movement and fMRI activity patterns for each sentence using ridge regression.
