Research map / Vision-language and multimodal / CLIP and zero-shot
MaPLe: Multi-Modal Prompt Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 coupled prompts in both vision and language branches instead of independent V-L prompts (ablation) | CLIP ViT-B/16 · Base-to-novel generalization, 11 datasets average | Harmonic mean: 77.90 -> 78.55 | Helped |
| Add language-branch deep prompts rather than only vision-branch deep prompts (ablation) | CLIP ViT-B/16 · Base-to-novel generalization, 11 datasets average | Harmonic mean: 76.68 -> 77.56 | Helped |
| Add deep prompts in both branches rather than shallow MaPLe prompts (ablation) | CLIP ViT-B/16 · Base-to-novel generalization, 11 datasets average | Harmonic mean: 76.67 -> 78.55 | Helped |
| Use one shared V-L coupling function across layers (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 78.11 | Hurt |
| Increase prompt length (ablation) | MaPLe · Base-to-novel generalization | Novel accuracy | Hurt |
| Initialize prompts in every layer with the same template rather than only the first layer (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 77.88 | Hurt |
| Randomly initialize prompts rather than use the template in the first layer (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 78.52 | Hurt |
| Project vision prompts to language prompts instead of language prompts to vision prompts (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 77.10 | Hurt |
| Add progressive conditioning of prompts across transformer blocks (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 78.11 | Hurt |
| Combine MaPLe with independent V-L prompting (ablation) | MaPLe · Base-to-novel generalization, 11 datasets average | Harmonic mean: 78.55 -> 77.94 | Hurt |
| Use multimodal branch-aware prompting instead of Co-CoOp | CLIP ViT-B/16 · Base-to-novel generalization, 11 datasets average | Novel accuracy: 71.69 -> 75.14 | Helped |
| Use multimodal branch-aware prompting instead of Co-CoOp | CLIP ViT-B/16 · Cross-dataset evaluation, 10 target datasets average | Accuracy: 65.74 -> 66.30 | Helped |
Closest papers by meaning
- Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.93)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.92)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (AAAI 2024 · similarity 0.92)
- Aggregate-and-Adapt Natural Language Prompts for Downstream Generalization of CLIP (NeurIPS 2024 · similarity 0.92)
- ViLT-CLIP: Video and Language Tuning CLIP with Multimodal Prompt Learning and Scenario-Guided Optimization (AAAI 2024 · similarity 0.92)
- COMMA: Co-articulated Multi-Modal Learning (AAAI 2024 · similarity 0.92)
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.