Research map / Vision-language and multimodal / Cross-modal learning
SDSA: Shallow-Deep Squeezing Adapter for Vision-Language Models
Hao Wang, Rui Zhu, Xiaoxu Li, Zhanyu Ma, Jing-Hao Xue
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 |
|---|---|---|---|
| use SDSA instead of MMA | CLIP ViT-B/16 · Base-to-novel average (11 datasets) | HM: 79.87 -> 80.96 | Improved |
| use SDSA instead of 2SFS | CLIP ViT-B/16 · Base-to-novel average (11 datasets) | Base accuracy: 85.55 -> 83.77 | Hurt |
| use SDSA instead of MMA | CLIP ViT-B/16 · Cross-dataset transfer from ImageNet (average) | accuracy: 66.61 -> 67.32 | Improved |
| add low-rank transformation and cross-attention together (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 71.86 -> 80.96 | Improved |
| apply SDSA to all layers 1-12 instead of 9-12 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 77.48 -> 80.96 | Improved |
| increase rank from 8 to 16 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 80.51 -> 80.96 | Improved |
| increase rank from 16 to 64 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | Novel accuracy: 78.34 -> 77.68 | Hurt |
| increase scaling factor alpha from 0.01 to 0.05 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 77.10 -> 80.96 | Improved |
| increase scaling factor beta from 0.01 to 0.10 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | Novel accuracy: 78.34 -> 76.54 | Hurt |
| increase scaling factor beta from 0.01 to 0.10 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | Base accuracy: 83.77 -> 84.06 | Improved |
| add random token masking at ratio 0.10 versus no masking (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 80.91 -> 80.96 | No clear effect |
| raise mask ratio from 0.10 to 0.30 (ablation) | SDSA · Base-to-novel ablation (11-dataset average, HM) | HM: 80.96 -> 80.63 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaPLe — MaPLe: Multi-Modal Prompt Learning (CVPR 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (ICLR 2025 · similarity 0.87)
- Deep Pre-Alignment for VLMs (ICML 2026 · similarity 0.86)
- MHA2MLA-VLM: Enabling DeepSeek’s Economical Multi-Head Latent Attention Across Vision-Language Models (AAAI 2026 · similarity 0.86)
- MoDA: Modulation Adapter for Fine-Grained Visual Understanding in Instructional MLLMs (ICML 2026 · similarity 0.86)
- DeepStack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs (NeurIPS 2024 · similarity 0.86)
- MMInference: Accelerating Pre-filling for Long-Context Visual Language Models via Modality-Aware Permutation Sparse Attention (ICML 2025 · similarity 0.85)
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.
