Research map / Vision-language and multimodal / CLIP and zero-shot
Transferable Model-agnostic Vision-Language Model Adaptation for Efficient Weak-to-Strong Generalization
Jihwan Park, Taehoon Song, Sanghyeok Lee, Miso Choi, Hyunwoo J. Kim
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 TransMiter adaptation transfer instead of target zero-shot | CLIP target VLMs · 11 datasets base-to-base | accuracy: RN50->ViT-B/16: 65.33 -> 77.16 | Improved |
| use TransMiter instead of EFT | adaptation transfer · 11 datasets base-to-base | accuracy / GFLOPs: e.g. 81.04 -> 82.05; 35.16 -> 0.01 GFLOPs | Improved |
| use TransMiter instead of Prompt Transfer | adaptation transfer · ViT-B/16 -> ViT-L/14 PromptSRC | accuracy: 73.41 -> 82.51 | Improved |
| use a stronger source model (ViT-B/16 instead of RN50) (ablation) | TransMiter with ViT-L/14 target · 11 datasets base-to-base | accuracy: 80.39 -> 82.05 | Improved |
| use PromptSRC instead of CoOp for source fine-tuning (ablation) | TransMiter · ViT-B/16 -> ViT-L/14 | accuracy: 82.05 -> 82.51 | Improved |
| use TransMiter instead of baselines on novel classes | target ViT-L/14 · base-to-novel average | novel accuracy: 80.85 (zero-shot) -> 81.07 | Mixed |
| add auxiliary class expansion to the prediction adapter (ablation) | TransMiter target · base-to-novel HM | HM: 80.09 -> 81.24 | Improved |
| add basis change to the prediction adapter (ablation) | TransMiter target · base-to-novel HM | HM: 80.09 -> 80.80 | Improved |
| combine auxiliary classes and basis change (ablation) | TransMiter target · base-to-base | accuracy: 76.25 -> 82.51 | Improved |
| fine-tune TransMiter with labeled data (TransMiter+) | supervised fine-tuning strategies · base-to-novel HM, ViT-L/14 | HM: PromptSRC 84.28 -> 84.69; MaPLe 82.35 -> 84.37 | Improved |
| use TransMiter+ with HPT/LwEIB strategy | HPT, LwEIB · base-to-novel HM | HM: HPT 84.95 -> 85.02; LwEIB 84.75 -> 84.96 | No clear effect |
| use TransMiter instead of Prompt Transfer for transfer cost | adaptation transfer · StanfordCars efficiency | transfer time (s): 480 -> 24 | Improved |
Closest papers by meaning
- MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · similarity 0.91)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.89)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation (NeurIPS 2024 · similarity 0.89)
- Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024 · 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.
