Cross-architecture, cross-task, and cross-subtype replication of the PRISM dissociation

Determine whether the phonemic-favoring-versus-semantic-favoring dissociation identified by the PRISM analysis replicates in pure-text language models, encoder-decoder architectures, language models at different parameter scales, additional clinical tasks such as the Western Aphasia Battery–Revised, sentence completion, and story comprehension, and additional aphasia subtypes including Broca’s, Wernicke’s, conduction, anomic, mixed transcortical, and global aphasia.

Background

The paper applies PRISM to a single vision-LLM, LLaVA-1.6-Vicuna-13B, and to the Philadelphia Naming Test in a cohort of chronic post-stroke aphasia patients. Its principal result is a robust phonemic-favoring dissociation relative to semantic errors across transformer layers and cortical lesion-symptom maps, while the semantic-favoring direction remains a consistently signed but non-significant trend. The authors explicitly identify as unresolved whether this dissociation generalizes beyond the particular architecture, scale, task, and patient population studied. Such replication would test whether the observed organization reflects a general property of language-model processing and aphasia-related language systems rather than a contingent feature of the selected model, behavioral measure, or clinical subtypes.