Research map / Vision-language and multimodal / CLIP and zero-shot
Test-Time Debiasing with Probabilistic Prompts via Wasserstein Distance in Vision-Language Models
Chengye Wang, Yuyuan Li, XiaoHua Feng, Xiaolin Zheng, Chaochao Chen
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 |
|---|---|---|---|
| replace point-based debiasing with Wasserstein distributional alignment plus probabilistic prompts | CLIP-ViT-B-P16 · UTKFace race, stereotype queries | MaxSkew: 1.0074 -> 0.3312 | Improved |
| replace point-based debiasing with Wasserstein distributional alignment plus probabilistic prompts | CLIP-ViT-B-P16 · UTKFace gender, stereotype queries | MaxSkew: 0.8204 -> 0.0833 | Improved |
| compare against BendVLM on binary gender attribute | CLIP-ViT-L-P14 · FairFace gender, stereotype queries | MaxSkew: 0.0952 (BendVLM) vs 0.1324 (W4D) | Hurt |
| increase beta (weight of performance preservation loss) from 0 to 2.0 (ablation) | W4D · CelebA/Facet zero-shot classification | Worst-Group AUC-ROC and MaxSkew | Mixed |
| increase gamma (orthogonality loss weight) from 0 to 3.0 (ablation) | W4D · CelebA/Facet zero-shot classification | Worst-Group AUC-ROC and MaxSkew | Mixed |
| increase Monte Carlo sample size K from 16 to 256 (ablation) | W4D probabilistic prompts · CelebA/Facet zero-shot classification | AUC and MaxSkew | Improved |
| apply orthogonal projection or orthogonal calibration debiasing | CLIP-ViT-B-P16 · FairFace gender, stereotype queries | MaxSkew: 0.6662 -> 1.3983 (Orth-Proj.) | Hurt |
| use W4D in zero-shot classification | CLIP · CelebA, Facet zero-shot classification | Worst-Group AUC-ROC | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BendVLM — BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024 · similarity 0.90)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.88)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.88)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.87)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · 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.
