Tag
This paper introduces PRISM, a perturbation-based method for spatially resolved interpretability of large language models, adapting neuroimaging subtraction analysis to transformers and applying it in parallel to post-stroke aphasia patients to recover shared phonemic-favoring dissociations.