Research map / Vision-language and multimodal / CLIP and zero-shot
A Closed-Form Solution for Debiasing Vision-Language Models with Utility Guarantees Across Modalities and Tasks
Tangzheng Lian, Guanyu Hu, Yijing Ren, Dimitrios Kollias, Oya Celiktutan
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 |
|---|---|---|---|
| use spherical mean of LLM-generated variants plus anchor as group prototype instead of anchor only (ablation) | Ours · Flickr30K MS@1000 (G) | MS@1000: 13.4 -> 11.8 | Improved |
| use anchor plus variants prototype instead of mean of variants only (ablation) | Ours · Flickr30K MS@1000 (G) | MS@1000: 14.1 -> 11.8 | Improved |
| debias both image and text instead of text only (ablation) | Ours · Flickr30K MS@1000 (G) | MS@1000: 13.3 -> 11.8 | Improved |
| debias both image and text instead of image only (ablation) | Ours · CelebA Delta Max EO (G x A) | Delta Max EO: 41.7 -> 40.1 | Improved |
| swap LLM for prototype generation (DeepSeek v3.2, Gemini 2.5 Pro vs GPT-5) (ablation) | Ours · Flickr30K MS@1000 / CelebA Delta Max EO | fairness metrics: MS 12.0 and 11.8 vs 11.8 | No clear effect |
| project orthogonally to attribute subspace with Pareto-optimal closed-form (Ours) instead of Orth-Proj | CLIP ViT-L/14 text-to-image retrieval · COCO2017 R@5 | R@5: 74.5 -> 81.1 | Improved |
| apply debiasing method to CLIP | CLIP ViT-L/14 · Flickr30K MS@1000 (G) | MS@1000: 20.3 -> 11.8 | Improved |
| apply debiasing method to CLIP | CLIP ViT-L/14 · CelebA zero-shot F1 | F1: 54.0 -> 56.5 | Improved |
| apply debiasing method to CLIP | CLIP ViT-L/14 · FACET Macro F1 | Macro F1: 70.8 -> 70.7 | No clear effect |
| apply debiasing to Stable Diffusion text embeddings | Stable Diffusion v2.1 · gender-explicit prompts AccG | AccG (%): 75.4 -> 74.6 | Mixed |
| debias only text embeddings (Orth-Proj) in SD | Stable Diffusion v2.1 · gender-explicit prompts AccG | AccG (%): 75.4 -> 53.4 | Hurt |
Closest papers by meaning
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.92)
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024 · similarity 0.91)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.91)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.91)
- FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication (CVPR 2024 · similarity 0.89)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
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.
