Research map / Vision-language and multimodal / Human and embodied modeling
Inducing Dyslexia in Vision Language Models
Melika Honarmand, Ayati Sharma, Badr AlKhamissi, Johannes Mehrer, Martin Schrimpf
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| ablate VWF-selective units instead of leaving the model intact (ablation) | Qwen2-VL-72B · ROAR | normalized accuracy: -32% | Hurt |
| ablate VWF-selective units (ablation) | Qwen2-VL-72B · RAVEN | accuracy: p≈0.75 | No clear effect |
| ablate VWF-selective units (ablation) | Qwen2-VL-72B · Kempler | accuracy: +10% | Improved |
| ablate random units in same layers instead of VWF-selective units (ablation) | Qwen2-VL-72B · ROAR, RAVEN, Kempler | accuracy: -21% ROAR, -21% RAVEN, -13% Kempler | Hurt |
| ablate VWF-selective units (ablation) | Qwen2-VL-72B · Phonology-sensitive lexical decision | accuracy: -8% | Hurt |
| ablate VWF-selective units (ablation) | Qwen2-VL-72B · Orthography-sensitive lexical decision | accuracy: p>0.059 | No clear effect |
| ablate units in vision encoder, merger or self-attention instead of decoder MLP (ablation) | Qwen2-VL-72B · ROAR vs RAVEN | selectivity | Hurt |
| scale VWF unit activity (factors -2 to 4) instead of full ablation (ablation) | Qwen2-VL-72B · ROAR | selective impairment | Hurt |
| use dyslexia-friendly fonts (OpenDyslexic, Comic Sans) with ablated model | ablated Qwen2-VL-72B · ROAR fonts | lexical decision accuracy: p << 0.001 vs Arial | Improved |
| use Papyrus font with ablated model | ablated Qwen2-VL-72B · ROAR fonts | lexical decision accuracy: p << 0.01 vs Arial | Hurt |
| apply VWF ablation to other VLMs | Molmo-72B, PixTral-12B · ROAR vs RAVEN | accuracy: Molmo 83% baseline | Improved |
Closest papers by meaning
- Artificial Aphasias in Lesioned Language Models (NeurIPS 2026 · similarity 0.86)
- In Silico Mapping of Visual Categorical Selectivity Across the Whole Brain (NeurIPS 2025 · similarity 0.82)
- fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.82)
- BrainSCUBA: Fine-Grained Natural Language Captions of Visual Cortex Selectivity (ICLR 2024 · similarity 0.82)
- Finding Shared Decodable Concepts and their Negations in the Brain (ICLR 2025 · similarity 0.82)
- Correlating instruction-tuning (in multimodal models) with vision-language processing (in the brain) (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
