Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Task-Adaptive Prompted Transformer for Cross-Domain Few-Shot Learning
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 MetaPrompt, a task-adaptive prompted ViT in which an attention-based prompt generator produces arbitrary-length task prompts from support features, combined with test-time bias tuning, for cross-domain few-shot learning on Meta-Dataset.
Problem
Cross-domain few-shot learning, where meta-test classes come from domains very different from meta-training, causing prior FSL methods to perform poorly.
Approach
Replace independent per-parameter generators (e.g., CNAPs FiLM generators) with an attention-based prompt generator (class-aware self-attention over support features plus task-aware cross-attention from learnable prompt tokens) that outputs a task-adaptive prompt of arbitrary length inserted into a ViT backbone.
Key results
- Achieves overall average accuracy of 86.0 on Meta-Dataset, a +2.2% gain over the prior best (PMF, 83.8).
- Out-of-domain average accuracy of 85.6 vs 82.0 for PMF (+3.1%), with gains of +4.6% on COCO and +4.9% on CF100.
- Ablation: prompt generator alone raises out-of-domain accuracy by 8.9% over baseline; adding bias tuning reaches 85.49 overall with the full combination (M6 in Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest baseline with MetaPrompt | Meta-Dataset · Meta-Dataset overall average | accuracy: 83.8 -> 86.0 | Helped |
| add the task-conditioned prompt generator (ablation) | MetaPrompt · Meta-Dataset | overall average accuracy: 76.71 -> 82.00 | Helped |
| add class-wise embedding (ablation) | MetaPrompt prompt generator · Meta-Dataset | overall average accuracy: 80.54 -> 82.00 | Helped |
| add class-aware self-attention (ablation) | MetaPrompt prompt generator · Meta-Dataset | overall average accuracy: 80.21 -> 82.00 | Helped |
| add bias tuning without the prompt generator (ablation) | MetaPrompt · Meta-Dataset | out-of-domain average accuracy: 66.31 -> 79.74 | Helped |
| add bias tuning to the prompt-generator method (ablation) | MetaPrompt · Meta-Dataset | overall average accuracy: 82.00 -> 85.49 | Helped |
| add bias tuning to the prompt-generator method (ablation) | MetaPrompt · Meta-Dataset | in-domain average accuracy: 86.21 -> 85.45 | Hurt |
| replace task-adaptive prompting with prompt tuning (lr=1) | MetaPrompt · Meta-Dataset overall average | accuracy: 86.0 -> 82.1 | Hurt |
| combine task-adaptive prompting with prompt tuning | MetaPrompt · Meta-Dataset overall average | accuracy: +2.8% | Helped |
| replace ViT-S (Dino) with ViT-B (Sup) (ablation) | MetaPrompt · Meta-Dataset overall average | accuracy: 86.0 -> 89.1 | Helped |
| replace attention-based prompt generation with average pooling and MLPs (ablation) | MetaPrompt · Meta-Dataset overall average | accuracy: 82.0 -> 79.5 | Hurt |
| increase prompt length from 8 to 12 (ablation) | MetaPrompt · Meta-Dataset overall average | accuracy: 82.0 -> 80.9 | Hurt |
Surprising findings
- Bias tuning hurts in-domain accuracy with a large learning rate, so it is applied only to out-of-domain datasets.
Limitations
- Evaluated only on Meta-Dataset image classification; other domains untested.
- Test-time bias tuning is required for best out-of-domain performance and is applied only to out-of-domain datasets.
Experimental setup
Datasets: Meta-Dataset MNS CF10 CF100
Benchmarks: Meta-Dataset
Metrics: accuracy in-domain average accuracy out-of-domain average accuracy overall average accuracy
Baselines: ProtoNet BOHB-E CNAPs SimpleCNAPs SUR URT FLUTE URL TSA PMF
Builds on: CNAPs SimpleCNAPs PMF TSA VPT
Closest papers by meaning
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 2024 · similarity 0.89)
- Random Registers for Cross-Domain Few-Shot Learning (ICML 2025 · similarity 0.88)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.87)
- Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach (CVPR 2024 · similarity 0.87)
- Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation (CVPR 2024 · similarity 0.86)
- Exploring Sparse Visual Prompt for Domain Adaptive Dense Prediction (AAAI 2024 · similarity 0.86)
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.