Reproduce “Inducing Dyslexia in Vision Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Melika Honarmand, Ayati Sharma, Badr AlKhamissi, Johannes Mehrer, Martin Schrimpf
Researchers from École Polytechnique Fédérale de Lausanne (EPFL) and the University of California, Berkeley developed a computational framework to induce dyslexia-like symptoms in vision-language models (VLMs) by functionally localizing and perturbing specific visual-word-form (VWF)-selective units. The resulting "dyslexic" model exhibited selective reading deficits, preserved general reasoning abilities, and accurately predicted the efficacy of dyslexia-friendly fonts.
Key insights
- Targeted perturbation of functionally localized VWF-selective units can selectively induce dyslexia-like reading impairments in VLMs while preserving general visual reasoning and syntactic comprehension.
- The ablated model demonstrated a specific phonological deficit, mirroring human dyslexic patterns, which provides computational support for the link between VWFA impairment and phonological dysfunction.
- Computational models can serve as a manipulable 'digital twin' for testing causal hypotheses about brain disorders and prototyping interventions like dyslexia-friendly fonts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.