Research map / Vision-language and multimodal / CLIP and zero-shot
SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCAP, a CLIP-based transductive test-time adaptation method that mines cross-sample visual cliques per test batch, learns visual and text attribute prompts per clique, and retains them via a graph-based cache.
Problem
Test-time adaptation of vision-language models under domain shift when test data arrives in batches; existing TTA methods treat samples individually or ignore the text modality in batch-level (transductive) adaptation.
Approach
Form supportive cliques of test images via visual similarity within a batch (from top-k class predictions), learn a visual and text attribute prompt per clique with entropy and concentration losses, aggregate prompts for each sample from its cliques, and maintain a retention cache with graph-based merging of attribute prompts across batches.
Key results
- On the OOD benchmark, SCAP reaches 65.63% average accuracy versus 63.89% for TDA and 63.48% for DART with CLIP ViT-B/16.
- SCAP reaches 64.52% on ImageNet-A, improving over the best prior method by at least 3.17 points.
- On cross-domain datasets SCAP improves over the best prior method by at least 1.64 points on Pets and about 0.2 points on Flower102, Caltech101, and Aircraft.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SCAP instead of TDA | CLIP ViT-B/16 · ImageNet-A | accuracy: 60.11 -> 64.52 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · ImageNet-R | accuracy: 80.24 -> 81.68 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · ImageNet-Sketch | accuracy: 50.54 -> 51.65 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · ImageNet-V2 | accuracy: 64.67 -> 64.65 | Hurt |
| apply SCAP instead of TDA | CLIP ViT-B/16 · ImageNet variants average | accuracy: 63.89 -> 65.63 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · Pets | accuracy: 88.63 -> 90.27 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · Flower102 | accuracy: 71.42 -> 71.65 | Helped |
| apply SCAP instead of TDA | CLIP ViT-B/16 · Caltech101 | accuracy: 94.24 -> 94.42 | Helped |
| apply SCAP instead of MTA | CLIP ViT-B/16 · Aircraft | accuracy: 25.20 -> 25.44 | Helped |
| add text attribute learning (ablation) | SCAP · ImageNet-A | accuracy: 63.14 -> 64.52 | Helped |
| add image attribute learning (ablation) | SCAP · ImageNet-A | accuracy: 62.57 -> 64.52 | Helped |
| add image attribute learning and visual retention (ablation) | SCAP · ImageNet-A | accuracy: 60.98 -> 63.82 | Helped |
Surprising findings
- On ImageNet-V2, SCAP performs on par with TDA (only a 0.02% difference), so gains there are marginal.
Limitations
- Performance is sensitive to the clique similarity threshold t; too low admits unrelated images and too high limits support samples.
- Gains on ImageNet-V2 and some cross-domain datasets are marginal.
Experimental setup
Datasets: ImageNet-A ImageNet-R ImageNet-Sketch ImageNet-V2 Aircraft Caltech101 Flower102 Pets
Benchmarks: ImageNet-A ImageNet-R ImageNet-Sketch ImageNet-V2 Aircraft Caltech101 Flower102 Pets
Metrics: accuracy (Acc@1)
Baselines: CLIP CoOp CoCoOp Tip-Adapter TPT C-TPT DART MTA TDA Zero CLIPArTT
Builds on: TPT DART TDA CLIP
Closest papers by meaning
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.90)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- Realistic Test-Time Adaptation of Vision-Language Models (CVPR 2025 · similarity 0.89)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.88)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · similarity 0.88)
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.