Research map / Vision-language and multimodal / CLIP and zero-shot
Unsupervised Vision-and-Language Pre-Training via Retrieval-Based Multi-Granular Alignment
Mingyang Zhou, Licheng Yu, Amanpreet Singh, Mengjiao Wang, Zhou Yu, Ning Zhang
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 |
|---|---|---|---|
| Use joint image-text input instead of alternating image-only and text-only input (ablation) | UVLP with MLM+MRM · V+L downstream tasks | meta-average: 67.4 -> 68.1 | Improved |
| Increase the proportion of originally aligned image-text pairs (ablation) | UVLP pre-training on CC · V+L downstream tasks | meta-average: 68.1 -> 73.5 -> 74.5 -> 76.0 -> 76.8 -> 77.3 | Improved |
| Add region-tag alignment to the model (ablation) | μ-VLA · V+L downstream tasks | meta-average: 73.7 -> 75.0 | Improved |
| Add region-noun phrase alignment to the model (ablation) | μ-VLA · V+L downstream tasks | meta-average: 69.5 -> 75.0 | Improved |
| Add image-sentence alignment to region-tag and region-noun phrase alignment (ablation) | μ-VLA · V+L downstream tasks | meta-average: 75.0 -> 75.8 | Improved |
| Use image-sentence alignment rather than region-tag alignment alone (ablation) | μ-VLA · NLVR2 Test-P | accuracy: 52.0 -> 71.5 | Improved |
| Use ITM alignment scores to weight retrieved-pair losses (ablation) | μ-VLA · V+L downstream tasks | meta-average: 75.3 -> 75.8 | Improved |
| Retrieve weakly aligned text instead of randomly pairing text and images (ablation) | μ-VLA on CC images and BookCorpus text · V+L downstream tasks | meta-average: 68.3 -> 73.8 | Improved |
| Increase retrieved candidates per image from 1 to 5 (ablation) | μ-VLA using image-sentence alignment · V+L downstream tasks | meta-average | Improved |
| Increase retrieved candidates per image from 5 to 10 (ablation) | μ-VLA using image-sentence alignment · V+L downstream tasks | meta-average | Hurt |
Closest papers by meaning
- Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models (ICML 2024 · similarity 0.90)
- Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning (NeurIPS 2026 · similarity 0.90)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.90)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.89)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.89)
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.89)
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.
