Research map / Vision-language and multimodal / Cross-modal learning
Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Shift embeddings to tune the modality gap | CLIP ViT-B/16 · CIFAR10, CIFAR100, EuroSAT, SVHN, HatefulMemes | zero-shot accuracy: CIFAR10: 0.9013 -> 0.9081; CIFAR100: 0.6658 -> 0.6737; EuroSAT: 0.5410 -> 0.5645; SVHN: 0.5389 -> 0.5396; HatefulMemes: | Mixed |
| Increase the modality gap from 0.82 to 0.97 | CLIP ViT-B/32 · FairFace | offensive top-1 predictions: Black: 1.1% -> 1.0%; White: 15.7% -> 13.7%; Indian: 1.2% -> 1.1%; Latino: 2.8% -> 2.0%; Middle Eastern: 6.3% -> 5.2%; So | Helped |
| Increase the gap from 0.82 to 0.97 | CLIP ViT-B/32 · FairFace | top-1 accuracy: 0.5817 to 0.5739 | Hurt |
| Reduce the modality gap to d=0.07 (ablation) | CLIP ViT-B/32 · FairFace | crime-related denigration: Black: 1.0% -> 2.3%; White: 15.5% -> 23.0%; Indian: 1.2% -> 3.2%; Latino: 2.8% -> 11.8%; Middle Eastern: 6.3% -> 16.7%; | Hurt |
| Increase the modality gap to d=1.29 (ablation) | CLIP ViT-B/32 · FairFace | non-human denigration: Black: 0.1% -> 40.5%; White: 0.2% -> 42.4%; Indian: 0.0% -> 5.1%; Latino: 0.1% -> 10.7%; Middle Eastern: 0.0% -> 18.9%; | Hurt |
| Align text and image embeddings at initialization (ablation) | CLIP trained from scratch · MSCOCO Caption | final modality gap distance: 1.2991 ± 0.0389 -> 0.7457 ± 0.0633 | Hurt |
| Increase contrastive-loss temperature (ablation) | CLIP · MSCOCO Caption | modality gap distance | Hurt |
| Remove mismatched pairs from the simulation (ablation) | Contrastive-learning simulation · 3D unit-sphere embedding simulation | repulsive loss-landscape structure | Hurt |
| Close the gap by manually shifting embeddings | CLIP at temperature τ=1/100 · MSCOCO Caption validation | contrastive loss: 0.82 | Hurt |
Closest papers by meaning
- Multi-modal contrastive learning adapts to intrinsic dimensions of shared latent variables (NeurIPS 2025 · similarity 0.89)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.89)
- Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models (ICLR 2025 · similarity 0.88)
- On the Comparison between Multi-modal and Single-modal Contrastive Learning (NeurIPS 2024 · similarity 0.88)
- Aligning Multimodal Representations through an Information Bottleneck (ICML 2025 · similarity 0.88)
- Accept the Modality Gap: An Exploration in the Hyperbolic Space (CVPR 2024 · similarity 0.88)
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.