Research map / Vision-language and multimodal / Cross-modal learning
Closing the Modality Gap Aligns Group-Wise Semantics
Eleonora Grassucci, Giordano Cicchetti, Emanuele Frasca, Aurelio Uncini, Danilo Comminiello
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 ATP and centroid uniformity losses to InfoNCE to close the modality gap | CLIP (LT) · MSCOCO (2 modal) | V-Measure: 12.98 -> 23.63 | Improved |
| add ATP and centroid uniformity losses to InfoNCE | CLIP (LT) · MSCOCO (2 modal) | text-image R@1: 74.6 -> 70.3 | Hurt |
| add ATP and centroid uniformity losses to InfoNCE | CLIP (LT) · MSR-VTT (3 modal) | V-Measure: 23.3 -> 32.1 | Improved |
| add ATP and centroid uniformity losses to InfoNCE | CLIP (FT) · MSR-VTT (3 modal) | text-video R@1: 34.8 -> 32.8 | Hurt |
| add the proposed gap-closing losses | CLIP · CIFAR10 (2 modal) | V-Measure: 67.0 -> 67.9 | Improved |
| add the proposed gap-closing losses | CLIP (LT) · AV-MNIST (3 modal) | V-Measure: 77.6 -> 82.7 | Improved |
| add the proposed gap-closing losses | CLIP (LT) · MSCOCO captioning | CIDEr: 153.2 -> 160.9 | Improved |
| add the proposed gap-closing losses | CoMM · MultiBench MOSI | accuracy: 65.91 -> 67.65 | Improved |
| add the proposed gap-closing losses | CoMM · MultiBench UR-FUNNY | accuracy: 62.83 -> 63.25 | Improved |
| add InfoNCE loss to ATP + CU losses (ablation) | gap-closing objective · MSR-VTT | T2V R@1: 30.0 -> 35.2 | Improved |
| add centroid uniformity (CU) weight lambda2 (ablation) | InfoNCE + ATP · CIFAR10 | V-Measure and R@1: R@1 79.39 -> 86.20, V 61.65 -> 66.64 (lambda1=0, lambda2 0 -> 1) | Improved |
| add ATP weight lambda1 on top of CU (ablation) | InfoNCE + CU · CIFAR10 | V-Measure: 66.64 -> 71.47 (lambda2=1.00, lambda1 0 -> 1.00) | Improved |
Closest papers by meaning
- Accept the Modality Gap: An Exploration in the Hyperbolic Space (CVPR 2024 · similarity 0.88)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 2024 · similarity 0.88)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.88)
- Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.88)
- Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion (ICLR 2025 · similarity 0.88)
- Test-Time Matching: Unlocking Compositional Reasoning in Multimodal Models (ICLR 2026 · 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.
