AlignedNorm: Prompting Vision–Language Models via Coupled Prompt Field
Published in International Conference on Machine Learning (ICML), 2026
AlignedNorm rethinks prompt learning for vision-language models as Coupled Prompt Field construction. By dynamically aligning prompt-token norms to the native scale of pretrained VLMs, it stabilizes task-agnostic coupling between base and new classes and improves generalization without external distillation or decoupled inference.
PDF | Project Page | 中译版 | Code | Official Version | Slides | BibTeX