Research map / Vision-language and multimodal / Cross-modal learning
Modal Aphasia: Can Unified Multimodal Models Describe Images From Memory?
Michael Aerni, Joshua Swanson, Kristina Nikolić, Florian Tramèr
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 |
|---|---|---|---|
| compare text description with image generation of the same movie posters | ChatGPT-5 · Movie posters from memory | rubric error rate: text 45% vs image 6% errors | Hurt |
| describe the learned faces in text instead of generating them | fine-tuned Janus-Pro and Harmon · Synthetic faces | attribute accuracy: description near random guessing (around 20%) | Hurt |
| test on unseen concept combinations in image generation (ablation) | fine-tuned Janus-Pro and Harmon · Abstract visual concepts (held-out test) | image generation accuracy | No clear effect |
| ask the model to visualize the poster before describing it | ChatGPT-5 · Movie posters from memory | description error rate | No clear effect |
| fine-tune only the LLM backbone, freezing other parameters | Janus-Pro and Harmon · Controlled synthetic experiments | modal aphasia emergence | No clear effect |
| text-only refusal training for feet prompts | Janus-Pro · Safety case study | refusal rate on rare expression: 89% (real word) -> 24% (rare expression) | Hurt |
| text-only refusal training | Janus-Pro · Safety case study | image generation capability for feet | No clear effect |
| different concept types (shape vs position) | Janus-Pro · Abstract visual concepts | verbal accuracy: shapes around 23% vs random 14%; positions below random 25% | Mixed |
Closest papers by meaning
- Quantifying Cross-Modality Memorization in Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.86)
- Robust Multimodal Large Language Models Against Modality Conflict (ICML 2025 · similarity 0.86)
- The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio (NeurIPS 2025 · similarity 0.86)
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.86)
- Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs (NeurIPS 2025 · similarity 0.85)
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.
