Researchers investigated whether controlled perturbations to a multimodal language model can reproduce the systematic naming errors characteristic of post-stroke aphasia. Using LLaVA 1.6, they applied noise to specific layers and units to simulate lesions and evaluated the results against data from 278 people with aphasia (PWAs).

  • The study classified responses into seven categories using a validated neural classifier on the Philadelphia Naming Test.
  • Six of the seven error types emerged at clinically-comparable proportions across distinct parameter regions, with formal paraphasia being the exception.
  • Searching the perturbation space identified configurations that reproduced individual error profiles in at least six of seven categories for 97.8% of PWAs and all seven for 79.5%.
  • Monte Carlo baselines confirmed that this matching reflects joint inter-category structure rather than marginal overlap.

These results establish a quantitative framework for reproducing individual aphasic error patterns, suggesting the potential for language models to serve as digital twins of individuals with post-stroke aphasia.