Researchers have developed a method to simulate aphasia-like naming errors in multimodal language models by introducing controlled perturbations. By applying lesions to LLaVA 1.6, they found that six out of seven common response categories in picture naming, such as semantic errors and unrelated responses, could be reproduced at clinically comparable proportions. The framework successfully matched the individual error profiles of 97.8% of aphasia patients in at least six categories and 79.5% in all seven, suggesting potential for language models to act as digital twins for individuals with post-stroke aphasia. AI
IMPACT This research could lead to new diagnostic tools for aphasia and a deeper understanding of language processing in both humans and AI.
RANK_REASON The cluster contains a research paper published on arXiv detailing a novel methodology and findings in AI.
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- LLaVA 1.6
- Philadelphia Naming Test
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
