Research map / Vision-language and multimodal / CLIP and zero-shot
VL-Match: Enhancing Vision-Language Pretraining with Token-Level and Instance-Level Matching
Junyu Bi, Daixuan Cheng, Ping Yao, Bochen Pang, Yuefeng Zhan, Chuanguang Yang, Yujing Wang, Hao Sun, Weiwei Deng, Qi Zhang
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 MLM to the text encoder (ITC + ITM baseline) (ablation) | VL-Match pretraining · NLVR2 + Flickr30k | Meta-sum: 340.25 -> 340.47 | No clear effect |
| add generator-discriminator structure (ablation) | VL-Match pretraining · NLVR2 + Flickr30k | Meta-sum: 340.47 -> 342.41 | Improved |
| replace MLM with VL-RTD in the multimodal encoder (ablation) | VL-Match pretraining · NLVR2 + Flickr30k | Meta-sum: 342.41 -> 344.37 | Improved |
| add FG-ITM with NegGen negatives (ablation) | VL-Match pretraining · NLVR2 + Flickr30k | Meta-sum: 344.37 -> 346.59 | Improved |
| mask nouns and adjectives in NegGen instead of random tokens (ablation) | NegGen · NLVR2 test-P | accuracy: 81.29 -> 82.23 | Improved |
| filter out generated texts that match the image in NegGen (ablation) | NegGen · NLVR2 test-P | accuracy: 81.34 -> 82.23 | Improved |
| add FG-ITM (ablation) | VL-Match · VALSE average | accuracy: 69.34 -> 71.22 | Improved |
| add FG-ITM (ablation) | VL-Match · SVO-Probes overall | accuracy: 87.96 -> 89.38 | Improved |
| set discriminator loss weight lambda to 10 instead of 1 (ablation) | VL-RTD · NLVR2 + Flickr30k | Meta-sum: 340.53 -> 344.37 | Improved |
| increase lambda beyond 10 (15, 20) (ablation) | VL-RTD · NLVR2 + Flickr30k | Meta-sum: 344.37 -> 342.74 (15), 343.63 (20) | Hurt |
| replace MLM with VL-RTD (ablation) | pretraining cost · pretraining FLOPs | FLOPs (1e19): 1.898 -> 1.952 | Hurt |
Closest papers by meaning
- ViLTA: Enhancing Vision-Language Pre-training through Textual Augmentation (ICCV 2023 · similarity 0.89)
- GRIT-VLP: Grouped Mini-Batch Sampling for Efficient Vision and Language Pre-training (ECCV 2022 · similarity 0.89)
- BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (ICML 2022 · similarity 0.89)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.88)
- Probing Visual Language Priors in VLMs (ICML 2025 · similarity 0.87)
- Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models (ICML 2024 · 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.
