Research map / Vision-language and multimodal / CLIP and zero-shot
POUF: Prompt-Oriented Unsupervised Fine-tuning for Large Pre-trained Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Adapt CLIP to unlabeled target images by aligning class text prototypes and image features | CLIP · Office-31, Office-Home, and DomainNet | accuracy (%): 77.1 -> 88.9; 82.4 -> 86.1; 57.7 -> 61.3 | Helped |
| Adapt RoBERTa-large using POUF on unlabeled text | RoBERTa-large · SST-2 | accuracy: 83.6 -> 89.4 (0.0) | Helped |
| Adapt RoBERTa-large using POUF on unlabeled text | RoBERTa-large · RTE | accuracy: 51.3 -> 64.5 (2.5) | Helped |
| Replace conditional transport with exact OT (ablation) | POUF · Office-31 and RTE | accuracy: 88.0 ± 0.9 -> 86.1 ± 0.2; 64.5 ± 2.5 -> 58.0 ± 5.5 | Hurt |
| Replace conditional transport with OT-Sinkhorn (ablation) | POUF · Office-31 and RTE | accuracy: 88.0 ± 0.9 -> 86.5± 0.2; 64.5 ± 2.5 -> 63.3± 6.0 | Hurt |
| Add transport loss (ablation) | POUF · Office-31 and RTE | accuracy: 83.3 ± 0.4 -> 88.0 ± 0.9; 56.9 ± 1.5 -> 64.5 ± 2.5 | Helped |
| Add mutual-information loss (ablation) | POUF · Office-31 and RTE | accuracy: 84.2 ± 0.2 -> 88.0 ± 0.9; 58.4 ± 1.3 -> 64.5 ± 2.5 | Helped |
| Replace cosine cost with an RBF-inspired cost (ablation) | POUF · Office-31 and RTE | accuracy: 87.6± 0.5 vs 88.0 ± 0.9; 61.3± 6.3 vs 64.5 ± 2.5 | Hurt |
| Tune the full model instead of only soft prompts (ablation) | POUF on CLIP · Office-31, Office-Home, and DomainNet | accuracy (%): 88.0 -> 88.9; 84.3 -> 86.1; 60.2 -> 61.3 | Helped |
| Use only seen-class prototypes rather than seen and unseen prototypes (ablation) | POUF on CLIP · Amazon in-domain unseen-class generalization | accuracy: 84.1 vs 80.6; 81.7 vs 87.2 | Mixed |
| Add POUF loss to LM-BFF using unlabeled few-shot examples | LM-BFF (man) · 14 of 15 language tasks | task performance | Helped |
| Vary the mutual-information weight λ (ablation) | POUF on CLIP · Office-31 Webcam | accuracy: 87.5; 88.3; 90.6; 91.7; 90.8; 89.2; 90.6; 89.8; 88.9; 91.6 | Mixed |
Closest papers by meaning
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · similarity 0.88)
- Consistency-guided Prompt Learning for Vision-Language Models (ICLR 2024 · similarity 0.88)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.87)
- Amend to Alignment: Decoupled Prompt Tuning for Mitigating Spurious Correlation in Vision-Language Models (ICML 2024 · similarity 0.87)
- ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models (ICLR 2025 · similarity 0.87)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.87)
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.