Research map / Vision-language and multimodal / Cross-modal learning
BridgeTower: Building Bridges between Encoders in Vision-Language Representation Learning
Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan
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 |
|---|---|---|---|
| Connect uni-modal encoder layers to each cross-modal layer with Add&Norm bridge layers | METER-style VL model · VQAv2 test-std | accuracy: 78.73% vs 77.64 | Improved |
| Connect uni-modal encoder layers to each cross-modal layer with bridge layers | METER-style VL model · Flickr30K 1K test | RSUM: 576.6 vs 570.7 | Improved |
| Replace Add&Norm with elementwise product for the bridge (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18 -> 73.41 | Hurt |
| Replace Add&Norm with learned interpolation, gated interpolation, or a residual learned projection (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18; 75.09; 75.13; 75.10 | No clear effect |
| Replace Add&Norm with concatenation projection, nonlinear projection, cross-attention, or cross-attention plus FFN (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18 -> 74.55; 74.26; 73.67; 73.54 | Hurt |
| Use 2, 3, 4, or 5 rather than 6 cross-modal layers (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18; 74.12; 74.36; 75.00; 74.98 | Hurt |
| Increase the cross-modal encoder from 6 to 8, 10, or 12 layers (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18; 75.07; 75.06; 74.94 | No clear effect |
| Replace internal bridge-connected layers with external cross-modal layers (ablation) | BridgeTower · VQAv2 test-dev | accuracy: 75.18 -> 74.04 | Hurt |
| Replace internal bridge-connected layers with weighted-sum multi-layer feature fusion (ablation) | BridgeTower · Flickr30K 1K test | RSUM: 533.8 -> 531.4 | Hurt |
| Add VG-QA COCO-only question-answer pairs during fine-tuning (ablation) | BridgeTower-base · VQAv2 test-dev | accuracy: 78.66 -> 78.77 | Improved |
| Add all VG-QA question-answer pairs during fine-tuning (ablation) | BridgeTower-base · VQAv2 test-dev | accuracy: 78.66 -> 79.10 | Improved |
Closest papers by meaning
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.90)
- MMRL: Multi-Modal Representation Learning for Vision-Language Models (CVPR 2025 · similarity 0.90)
- Unveiling Encoder-Free Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Florence-VL: Enhancing Vision-Language Models with Generative Vision Encoder and Depth-Breadth Fusion (CVPR 2025 · similarity 0.89)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · 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.
