Research map / Vision-language and multimodal / CLIP and zero-shot
Debiased Fine-Tuning for Vision-Language Models by Prompt Regularization
Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, Hanwang Zhang
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 |
|---|---|---|---|
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · BAR | accuracy: 90.9 -> 92.1 | Improved |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · DomainNet OOD | accuracy: 63.7 -> 64.4 | Improved |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · DomainNet ID | accuracy: 75.9 -> 75.6 | Hurt |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · DomainNet | harmonic mean: 67.4 -> 69.2 | Improved |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · NICO Animal OOD | accuracy: 92.5 -> 94.4 | Improved |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · NICO Vehicle OOD | accuracy: 95.9 -> 94.1 | Hurt |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | CLIP fine-tuning · PACS | average accuracy: 94.7 -> 96.0 | Improved |
| Regularize fine-tuning with zero-shot prompt predictions using a sample-wise adaptive weight | ViLT fine-tuning · VQA-CP | accuracy: 46.34 -> 54.89 | Improved |
| Use sample-wise adaptive weighting instead of a constant distillation weight (ablation) | ProReg fine-tuning · VQA-CP | accuracy: 53.15; -1.74% gap to ProReg fine-tune | Improved |
| Use sample-wise adaptive weighting instead of a constant distillation weight (ablation) | ProReg fine-tuning · NICO Vehicle | harmonic mean | Improved |
| Ensemble zero-shot and conventional fine-tuning predictions instead of using ProReg (ablation) | CLIP and ViLT predictions · VQA-CP | accuracy: 53.48% vs 54.89% | Hurt |
| Ensemble zero-shot and conventional fine-tuning predictions instead of using ProReg (ablation) | CLIP fine-tuning · NICO Vehicle | harmonic mean | Hurt |
Closest papers by meaning
- Prompt Learning via Meta-Regularization (CVPR 2024 · similarity 0.91)
- Prompt-Robust Vision-Language Models via Meta-Finetuning (ICLR 2026 · similarity 0.91)
- Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models (ICML 2025 · similarity 0.90)
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · similarity 0.90)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.89)
- Consistency-guided Prompt Learning for Vision-Language Models (ICLR 2024 · 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.
