Research·arXiv cs.CL·14h ago

Lesioned language models reproduce stroke aphasia naming errors

Researchers demonstrated that systematically damaging multimodal language models can reproduce the specific error patterns observed in stroke-induced aphasia, validating LLMs as computational proxies for human language pathology. By applying targeted perturbations to LLaVA 1.6 across different layers and noise levels, the team successfully replicated six of seven error categories from the Philadelphia Naming Test across 278 patients with aphasia. This work bridges neuroscience and AI, suggesting that language model internals may reflect principles of human linguistic organization and opening pathways for clinical simulation, diagnostic modeling, and interpretability research grounded in real neurological data.

Modelwire context

Explainer

The key move here is methodological inversion: rather than asking whether LLMs behave like healthy humans, the researchers asked whether damaged LLMs behave like damaged humans. That framing is what makes the validation meaningful, because aphasia error categories are clinically precise and independently established, giving the model's failures a real external benchmark to fail against.

This connects directly to the route-choice bias paper from the same day ('Reproducing human biases in route choice using large language models'), which also treats LLMs as behavioral proxies, calibrated against known human patterns rather than abstract benchmarks. Both papers are part of a quiet methodological shift: using foundation models not to exceed human performance but to simulate human limitation and variability at scale. The aphasia work goes further by grounding that simulation in neurological data from 278 real patients, which raises the evidentiary bar considerably compared to survey-calibrated behavioral models.

The critical next test is whether the same lesioning approach generalizes beyond picture-naming to syntactic or phonological aphasia subtypes. If it does, and if the error-category match holds across languages other than English, the clinical simulation case becomes substantially harder to dismiss.

Coverage we drew on

Reproducing human biases in route choice using large language models: Toward scalable behavioral modeling · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsLLaVA 1.6 · Philadelphia Naming Test

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Lesioned Multimodal Language Models Reproduce Aphasic Picture-Naming Patterns”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.