Interpreting Brain Responses to Language with Sparse Features from Language Models

In three linesResearchers use sparse autoencoders (SAE) from language models to interpret brain responses to language via 7T fMRI. Testing 8 participants listening to 200 sentences, they identify voxel populations tuned to people-related content and show frontal regions are explained by surprisal alone, while the fronto-temporal network shares common features across regions.

Summary generated by Claude — human-verified