Research map / Vision-language and multimodal / CLIP and zero-shot
Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models
Dachuan Zhao, Weiyue Li, Zhenda Shen, Yushu Qiu, Bowen Xu, Haoyu Chen, Yongchao 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 SFID coordinate imputation with SPD subspace projection | SFID · FACET zero-shot, CLIP ViT-B/32 | Delta DP: 10.15 -> 9.94 | Improved |
| Replace SFID with SPD for retrieval | SFID · Flickr30K retrieval, CLIP RN50 | Skew@100: 0.1414 -> 0.1177 | Improved |
| Replace SFID with SPD for retrieval | SFID · Flickr30K retrieval, XVLM | Skew@100: 0.2032 -> 0.1859 | Improved |
| Replace SFID with SPD on classification accuracy | SFID · FACET zero-shot, CLIP ViT-B/32 | top-1 accuracy (%): 52.14 -> 51.29 | Hurt |
| Replace SFID with SPD in SDXL image generation | SFID · SDXL gender-prompt composite mismatch | composite mismatch rate: 1.74 -> 1.67 | Improved |
| Remove 5 subspace directions (INLP) instead of replacing 100 coordinates | SFID · FairFace gender probe, CLIP ViT-B/32 | probe accuracy: 0.9404 -> 0.6766 | Improved |
| Remove only 1 subspace direction (ablation) | SPD · FairFace gender probe, CLIP ViT-B/32 | probe accuracy: 0.9466 -> 0.9286 | No clear effect |
| Increase removed directions from 5 to 10 (ablation) | SPD · FairFace, non-target attribute probes | probe accuracy | Hurt |
| Add neutral mean reinjection after projection (ablation) | SPD · FACET zero-shot, CLIP RN50 | top-1 accuracy (%): 50.16 -> 51.44 | Improved |
| Add neutral mean reinjection after projection (ablation) | SPD · FACET zero-shot, XVLM | Delta DP: 9.87 -> 9.85 | No clear effect |
| Raise reinjection threshold tau from 0.7 to 0.9 (ablation) | SPD · FACET zero-shot, XVLM | Delta DP: 9.85 -> 10.26 | Hurt |
| Use adversarial residual DeAR in SDXL generation | DeAR · SDXL gender-prompt composite mismatch | composite mismatch rate: 4.42 -> 99.81 | Hurt |
Closest papers by meaning
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.91)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.90)
- 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 0.90)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.88)
- GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models (ICML 2025 · 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.
