Research map / Vision-language and multimodal / Cross-modal learning
Quantifying Cross-Modality Memorization in Vision-Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies cross-modality factual memorization in VLMs by fine-tuning Gemma-3-4B on a synthetic persona dataset (images and descriptions), finding transfer with a persistent source-target gap that image-caption augmentation partly closes.
Problem
How factual knowledge memorized in one modality (image or text) transfers to the other in vision-language models, and whether this transfer is consistent across training, scale, unlearning, and multi-hop settings.
Approach
Uses a controlled synthetic persona dataset (generated images and descriptions linked to names and favorite numbers) to train a VLM on one modality and test recall in the other, quantifying a cross-modal transferability gap and probing unlearning, multi-hop, and unintentional PII memorization.
Key results
- Facts learned in one modality transfer to the other, but a significant gap remains between source and target modality accuracy for both training settings.
- Increasing model size from 4B to 27B improves within-modality accuracy but leaves the cross-modal transfer slope nearly unchanged.
- Adding in-distribution synthetic image-caption augmentation improves cross-modal transferability, while COCO out-of-distribution data does not; in multi-hop tests, Desc=>Name plus Name=>FN training yields only 11% Desc=>FN accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Augment training with in-distribution synthetic image-text pairs | Gemma-3-4B · Persona image↔description recognition | cross-modal transferability | Helped |
| Augment training with out-of-distribution COCO images and captions | Gemma-3-4B · Persona image↔description recognition | cross-modal transferability | No clear effect |
| Train with varied image or description inputs rather than a single repeated input | Gemma-3-4B · Persona recognition | test accuracy | Helped |
| Train with varied inputs rather than a single repeated input | Gemma-3-4B · Persona image↔description recognition | cross-modal transferability rate | No clear effect |
| Increase model size from Gemma3-4B-IT to larger Gemma3 variants | Gemma3 family · Persona recognition | accuracy in source and target modalities | Helped |
| Increase model size from Gemma3-4B-IT to larger Gemma3 variants | Gemma3 family · Persona image↔description recognition | cross-modal transferability rate | No clear effect |
| Use image training rather than description training | Gemma-3-4B · Persona name recall across modalities | cross-modal transferability | Helped |
| Use a prompt asking the model to recall knowledge from another domain | Gemma-3-4B · Persona image↔description recognition | cross-modal transferability | No clear effect |
| Use a prompt asking the model to consider what it learned from the other modality | Gemma-3-4B · Persona image↔description recognition | cross-modal transferability | No clear effect |
| Train on Desc⇒Name and Name⇒FN, then query Desc⇒FN | Gemma-3-4B cross-modal multi-hop model · Desc⇒FN test | accuracy: 1.00 -> 0.11 | Hurt |
| Train with Image⇒Name base data instead of Desc⇒Name base data | Gemma-3-4B cross-modal multi-hop model · Image⇒FN test | accuracy | Helped |
| Train with Image⇒Name base data instead of Desc⇒Name base data | Gemma-3-4B cross-modal multi-hop model · Desc⇒FN test | accuracy | Helped |
Surprising findings
- Membership inference reveals the model memorizes an irrelevant favorite number embedded in training images (AUC around 70% after 100 epochs), while direct extraction succeeds near zero percent of the time.
- Naive prompts that ask the model to recall knowledge from another modality do not significantly change transferability.
Limitations
- Only text and image modalities and a single model family (Gemma 3) are studied.
- Synthetic persona setting with 100 personas may not reflect real-world knowledge acquisition.
- Unlearning is simulated by relabeling with different names rather than standard gradient-based unlearning, which the authors found unstable.
Experimental setup
Datasets: synthetic persona dataset COCO
Benchmarks: synthetic persona dataset
Metrics: recognition accuracy cross-modality accuracy unlearning accuracy drop membership inference AUC
Baselines: Gemma3-12B-IT Gemma3-27B-IT
Builds on: LoRA Gemma 3
Closest papers by meaning
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.89)
- Learning Modality Knowledge Alignment for Cross-Modality Transfer (ICML 2024 · similarity 0.88)
- Déjà Vu Memorization in Vision–Language Models (NeurIPS 2024 · similarity 0.88)
- Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval (NeurIPS 2025 · similarity 0.88)
- Towards General Continuous Memory for Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Captured by Captions: On Memorization and its Mitigation in CLIP Models (ICLR 2025 · similarity 0.87)
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.