Research map / Vision-language and multimodal / Multimodal LLMs
Task Bias in Contrastive Vision-Language Models
Sachit Menon, Ishaan Preetam Chandratreya, Carl Vondrick
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 |
|---|---|---|---|
| Add manual task-directed text prefix to labels (text vs actions) | CLIP zero-shot · Scene text vs Actions | proportion choosing intended task: +14.7% for text, 7.28% for actions | Mixed |
| Add manual task-directed text prefix to labels (objects vs text) | CLIP zero-shot · Objects vs Scene text | proportion choosing intended task: 9.22% decrease for objects, 11% increase for text | Mixed |
| Add manual task-directed text prefix to labels (objects vs actions) | CLIP zero-shot · Objects vs Actions | proportion choosing intended task: 37.5% decrease for objects, 29.6% increase for actions | Mixed |
| Add learned visual token prompt (ViTP) to CLIP | CLIP · Objects vs Text (target: Objects) | task preference accuracy (%): 50.28 -> 99.46 | Improved |
| Add learned visual token prompt (ViTP) to CLIP | CLIP · Objects vs Actions (target: Objects) | task preference accuracy (%): 48.68 -> 100.00 | Improved |
| Add learned visual token prompt (ViTP) to CLIP | CLIP · Actions vs Text (target: Actions) | task preference accuracy (%): 43.70 -> 100.00 | Improved |
| Increase pixel prompt border size from 1 to 5 (ablation) | Visual prompt (VP) · Objects vs Actions (target: Objects) | task preference accuracy (%): 67.55 -> 81.13 | Improved |
| Increase pixel prompt border size from 1 to 5 (ablation) | Visual prompt (VP) · OpenImages object recognition | accuracy (%): 37.254 -> 20.799 | Hurt |
| Add object-directed visual prompt to CLIP (ViTP) | CLIP · OpenImages object recognition | accuracy (%): 20.467 -> 27.547 | Improved |
| Add object-directed pixel prompt with border size 1 | CLIP · OpenImages object recognition | accuracy (%): 20.467 -> 37.254 | Improved |
| Use CLIP embedding instead of raw image to predict task bias direction (ablation) | Task bias classifier · Objects v. Scene Text | test accuracy (%): 61.4 -> 71.7 | Improved |
| Add attention overlay to image input of the task bias classifier (ablation) | Image classifier · Objects v. Scene Text | test accuracy (%): 61.4 -> 62.6 | Improved |
Closest papers by meaning
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.88)
- Prompt Learning via Meta-Regularization (CVPR 2024 · similarity 0.88)
- Aggregate-and-Adapt Natural Language Prompts for Downstream Generalization of CLIP (NeurIPS 2024 · similarity 0.88)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (AAAI 2024 · similarity 0.88)
- LaViP: Language-Grounded Visual Prompting (AAAI 2024 · 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.
