Research map / Vision-language and multimodal / CLIP and zero-shot
PRISM: Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection
Mahdiyar Molahasani, Azadeh Motamedi, Michael Greenspan, Il-Min Kim, Ali Etemad
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 orthogonal projection (PRISM-mini) with learned projection trained by LD loss (ablation) | PRISM · Waterbirds | worst-group accuracy: 69.5% -> 84.2% | Improved |
| Replace orthogonal projection (PRISM-mini) with learned projection trained by LD loss (ablation) | PRISM · CelebA | worst-group accuracy: 82.6% -> 84.0% | Improved |
| Apply PRISM debiasing projection to zero-shot CLIP-ViT-L/14 | CLIP-ViT-L/14 · Waterbirds | worst-group accuracy: 36.4% -> 84.2% | Improved |
| Apply PRISM debiasing projection to zero-shot CLIP-ViT-L/14 | CLIP-ViT-L/14 · CelebA | worst-group accuracy: 72.8% -> 84.0% | Improved |
| Apply PRISM to CLIP-RN50 | CLIP-RN50 · Waterbirds | worst-group accuracy: 38.3% -> 70.2% | Improved |
| Apply PRISM to CLIP-RN50 | CLIP-RN50 · CelebA | worst-group accuracy: 75.8% -> 83.3% | Improved |
| Use GPT 4o instead of Llama 3.2 (70B) as the LLM (ablation) | PRISM · CelebA | worst-group accuracy: 58.9% -> 84.0% | Improved |
| Use GPT 4o instead of GPT 4o mini as the LLM (ablation) | PRISM · CelebA | worst-group accuracy: 81.6% -> 84.0% | Improved |
| Use GPT 4o instead of GPT o1 as the LLM (ablation) | PRISM · CelebA | worst-group accuracy: 77.9% -> 84.0% | Improved |
| Use PRISM instead of Orth-Cali on CLIP-RN50 | PRISM · Waterbirds (CLIP-RN50) | worst-group accuracy: 74.0% -> 70.2% | Hurt |
| Evaluate scene descriptions in text instead of images (bias similarity) | CLIP zero-shot · Waterbirds | worst-group accuracy: 38.3% -> 33.1% | No clear effect |
| Set margin m to 0.6 in LD loss (ablation) | PRISM · Waterbirds and CelebA | worst-group accuracy | Improved |
Closest papers by meaning
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024 · similarity 0.90)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.90)
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.89)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.88)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.88)
- VLM-RobustBench: A Comprehensive Benchmark for Robustness of 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.
