Lesioned Multimodal Language Models Reproduce Aphasic Picture-Naming Patterns

This research explores whether general-purpose artificial intelligence can simulate the specific language impairments seen in patients with aphasia following a stroke. By applying controlled "lesions"—or targeted disruptions—to a multimodal language model, the researchers investigated if they could replicate the distinct patterns of naming errors that human patients experience. The study suggests that these models could potentially function as "digital twins," providing a new way to understand and model individual clinical profiles.

Testing the Model

The researchers utilized LLaVA 1.6, a multimodal language model, to see if it could mimic human performance on the Philadelphia Naming Test. To simulate the effects of brain damage, the team introduced perturbations to the model by varying which layers were affected, the proportion of units disrupted, and the intensity of the noise applied. They compared these simulated results against data from 278 individuals with aphasia, using a neural classifier to categorize naming responses into seven specific types, such as semantic errors, neologisms, or correct responses.

Key Findings

The study found that the model successfully reproduced six of the seven error categories at rates comparable to those observed in clinical settings. The only exception was "formal paraphasia." When the researchers searched through the model's parameter space, they were able to find specific configurations that matched the unique error profiles of individual patients. For 97.8% of the participants, the model matched at least six out of seven error categories, and for 79.5% of participants, it matched all seven. Statistical testing confirmed that these matches were not due to chance, but rather reflected the underlying structure of the error patterns.

Implications for Clinical Research

These results establish a quantitative framework for using AI to replicate complex, individual-specific language deficits. By demonstrating that a general-purpose model can be "lesioned" to mirror the specific naming errors of a person with aphasia, the researchers highlight the potential for these models to serve as digital twins. This approach could offer a new, scalable way to study post-stroke language recovery and provide deeper insights into the nature of aphasic naming patterns.