Research map / Vision-language and multimodal / CLIP and zero-shot
Personalize Your Large Vision-language Models With In-context Prompt Tuning
Yanshu Li, Jiaqian Li, Kuai Yu, Xi Xiao, Dongfang Liu, Tianyang Wang, Ruixiang Tang
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 the LVLM's native vision encoder instead of CLIP (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.806 -> 0.868 | Improved |
| use the LVLM's native vision encoder instead of DINO-v2 (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.825 -> 0.868 | Improved |
| use the LVLM's native vision encoder instead of DINO-v3 (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.846 -> 0.868 | Improved |
| fuse early and deep layer features instead of early only (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.824 -> 0.868 | Improved |
| fuse early and deep layer features instead of late only (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.817 -> 0.868 | Improved |
| use layers 3 and third-to-last instead of first and fifth from both ends (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.847 -> 0.868 | Improved |
| use alpha=0.3 for L_Geo instead of 0.1 (ablation) | ICPT (LLaVA-NeXT-7B) · Captioning (weighted) | concept recall + embedding similarity: 0.606 -> 0.654 | Improved |
| use alpha=0.3 instead of alpha=0.5 (ablation) | ICPT (LLaVA-NeXT-7B) · Recognition (weighted) | recall: 0.866 -> 0.868 | No clear effect |
| add Dynamic Token Router instead of random pruning at same ratio (ablation) | ICPT (LLaVA-NeXT-7B) · prompt-length sweep | personalization performance | Improved |
| apply MCS to both label embeddings and prompts instead of one only (ablation) | ICPT (LLaVA-NeXT-7B) · multi-concept tasks | personalization performance | Improved |
| use margin m=0.15 instead of strict orthogonality m=0 (ablation) | ICPT (LLaVA-NeXT-7B) · multi-concept tasks | personalization performance | Improved |
| add CVM and MCS with L_Geo (ablation) | ICPT (LLaVA-NeXT-7B) · single and multi-concept tasks | personalization performance | Improved |
Closest papers by meaning
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (AAAI 2024 · similarity 0.90)
- CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.90)
- ViP-LLaVA: Making Large Multimodal Models Understand Arbitrary Visual Prompts (CVPR 2024 · similarity 0.89)
- Exploring the Transferability of Visual Prompting for Multimodal Large Language Models (CVPR 2024 · 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.
