Research map / Vision-language and multimodal / CLIP and zero-shot
Uncertainty-guided Compositional Alignment with Part-to-Whole Semantic Representativeness in Hyperbolic Vision-Language Models
Hayeon Kim, Ji Ha Jang, Junghun James Kim, Se Young Chun
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 uncertainty calibration in the entailment loss (ablation) | UNCHA · zero-shot classification (general datasets), ViT-S/16 | accuracy: 64.57 -> 68.98 | Improved |
| add uncertainty-aware temperature scaling in global-local contrastive loss (ablation) | UNCHA · zero-shot classification (general datasets), ViT-S/16 | accuracy: 65.14 -> 68.98 | Improved |
| add entropy regularizer to uncertainty calibration (ablation) | UNCHA · zero-shot classification (general datasets), ViT-S/16 | accuracy: 65.61 -> 68.98 | Improved |
| add uncertainty calibration in the entailment loss (ablation) | UNCHA · image retrieval (Table 4), ViT-S/16 | retrieval score: 69.68 -> 73.90 | Improved |
| add entropy regularizer to uncertainty calibration (ablation) | UNCHA · misc. datasets classification, ViT-S/16 | accuracy: 24.78 -> 27.55 | Improved |
| uncertainty-guided part-to-whole alignment vs HyCoCLIP | hyperbolic VLM (ViT-B/16) · ImageNet zero-shot | top-1 accuracy: 45.8 -> 48.8 | Improved |
| uncertainty-guided part-to-whole alignment vs HyCoCLIP | hyperbolic VLM (ViT-B/16) · fine-grained datasets (CUB, Cars, Aircraft, Pets, Flowers, DTD) | top-1 accuracy: CUB 16.4 -> 14.8, Cars 11.6 -> 14.0 | Mixed |
| uncertainty-guided part-to-whole alignment vs HyCoCLIP | hyperbolic VLM (ViT-S/16) · EuroSAT zero-shot | top-1 accuracy: 38.3 -> 36.2 | Hurt |
| uncertainty-guided alignment vs HyCoCLIP on retrieval | hyperbolic VLM (ViT-B/16) · Flickr text retrieval R@1 | R@1: 92.6 -> 91.4 | Hurt |
| uncertainty-guided alignment vs HyCoCLIP on hierarchy metrics | hyperbolic VLM (ViT-B/16) · ImageNet hierarchical classification | TIE: 3.17 -> 2.94 | Improved |
| uncertainty-guided alignment vs HyCoCLIP on multi-object scenes | hyperbolic VLM (ViT-B/16) · ComCo 5 objects | mAP: 72.90 -> 81.18 | Improved |
| uncertainty-guided alignment vs HyCoCLIP on hard negatives | hyperbolic VLM (ViT-B/16) · Densely Captioned Images, All Hard Negs | accuracy: 53.33 -> 56.51 | Improved |
Closest papers by meaning
- Compositional Entailment Learning for Hyperbolic Vision-Language Models (ICLR 2025 · similarity 0.92)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · similarity 0.89)
- HyperET: Efficient Training in Hyperbolic Space for Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.88)
- Exploiting the Asymmetric Uncertainty Structure of Pre-trained VLMs on the Unit Hypersphere (NeurIPS 2025 · similarity 0.88)
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · similarity 0.87)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (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.
