Research map / Vision-language and multimodal / CLIP and zero-shot
Parts of Speech–Grounded Subspaces in Vision-Language Models
James Oldfield, Christos Tzelepis, Yannis Panagakis, Mihalis Nicolaou, Ioannis Patras
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 |
|---|---|---|---|
| project image embedding onto noun submanifold | CLIP ViT-B-32 · ImageNet | top-1 zero-shot accuracy: 56.50 -> 57.80 | Improved |
| project image embedding onto noun submanifold | CLIP ViT-B-32 · UT Zappos | top-1 zero-shot accuracy: 82.20 -> 87.70 | Improved |
| project image embedding onto noun submanifold | CLIP ViT-B-32 · FER2013 | top-1 zero-shot accuracy: 37.80 -> 44.70 | Improved |
| project image embedding onto noun submanifold | CLIP ViT-B-32 · Flowers102 | top-1 zero-shot accuracy: 62.50 -> 61.00 | Hurt |
| project image embedding onto noun submanifold | CLIP ViT-B-16 · MNIST | top-1 zero-shot accuracy: 48.00 -> 55.00 | Improved |
| project image embedding onto noun submanifold | CLIP ViT-B-16 · UT Zappos | top-1 zero-shot accuracy: 89.10 -> 88.10 | Hurt |
| project image embedding onto noun submanifold | CLIP ViT-L-14 · CIFAR10 | top-1 zero-shot accuracy: 91.70 -> 90.60 | Hurt |
| project image embedding onto noun submanifold | CLIP ViT-L-14 · FER2013 | top-1 zero-shot accuracy: 35.70 -> 46.40 | Improved |
| project image embedding onto noun submanifold | CLIP ViT-L-14 · MNIST | top-1 zero-shot accuracy: 59.70 -> 58.90 | Hurt |
| use geometry-aware submanifold instead of Euclidean PoS PGA subspace (ablation) | noun projection · ImageNet | top-1 zero-shot accuracy: 56.70 -> 57.80 | Improved |
| use Euclidean PoS PCA subspace instead of CLIP baseline (ablation) | CLIP ViT-B-32 · ImageNet | top-1 zero-shot accuracy: 56.50 -> 56.50 | No clear effect |
| lambda 0 vs 0.5 (no negative PoS guidance) (ablation) | PoS subspace · class invariance (qualitative) | top-1 zero-shot accuracy | Mixed |
Closest papers by meaning
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.88)
- Diffusion Feedback Helps CLIP See Better (ICLR 2025 · similarity 0.88)
- VITRIX-CLIPIN: Enhancing Fine-Grained Visual Understanding in CLIP via Instruction-Editing Data and Long Captions (NeurIPS 2025 · similarity 0.88)
- Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIP (NeurIPS 2024 · similarity 0.88)
- Object-centric binding in Contrastive Language-Image Pretraining (NeurIPS 2025 · similarity 0.87)
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 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.
