Research map / Vision-language and multimodal / Cross-modal learning
Mitigating the Modality Gap in Vision–Language Models with Fractal Spectral Geometry
Zihan Zhou, Yang Zhou, Ruoming Jin, Pan He, Patrick Emami
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 |
|---|---|---|---|
| add local zeta matching on top of the fractal constraint (ablation) | FSAlign · MS-COCO image-to-text (CLIP) | Recall@1: 45.56 -> 51.26 | Improved |
| add the multi-scale fractal constraint on top of zeta matching (ablation) | FSAlign · MS-COCO image-to-text (CLIP) | Recall@1: 43.18 -> 51.26 | Improved |
| add both fractal and zeta terms (ablation) | FSAlign · Tiny ImageNet-200 zero-shot (CLIP) | top-1 accuracy (%): 47.81 / 49.20 -> 51.85 | Improved |
| apply FSAlign geometry layer to frozen CLIP | CLIP · Flickr30k image-to-text | Recall@1: 69.40 -> 78.30 | Improved |
| apply FSAlign geometry layer to frozen CLIP | CLIP · Tiny ImageNet-200 zero-shot | top-1 accuracy (%): 40.31 -> 51.85 | Improved |
| apply FSAlign geometry layer to frozen CLIP | CLIP · MS-COCO modality gap | CMAS: 0.2974 -> 0.6490 | Improved |
| apply FSAlign to CLIP | CLIP · MS-COCO modality gap | centroid distance (CD): 0.8876 -> 0.7729 | Mixed |
| use gap-reducing baselines MG and AlignCLIP instead of plain CLIP | CLIP · CIFAR-100 zero-shot | top-1 accuracy (%): 62.62 -> 53.12 (MG), 52.09 (AlignCLIP) | Hurt |
| add FSAlign fractal constraint | CLIP embeddings · ball-mass power-law fit | R2 (image/text): 0.980/0.945 -> 0.996/0.993 | Improved |
Closest papers by meaning
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · similarity 0.88)
- Do Vision and Language Encoders Represent the World Similarly? (CVPR 2024 · similarity 0.87)
- Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment (NeurIPS 2024 · similarity 0.87)
- Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning (NeurIPS 2026 · similarity 0.87)
- Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling (NeurIPS 2026 · similarity 0.87)
- Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (NeurIPS 2026 · 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.
