Enhancing Continual Learning of Vision-Language Models via Dynamic Prefix Weighting [CVPR 2026]

This repository contains the official implementation of our CVPR 2026 paper:

Enhancing Continual Learning of Vision-Language Models via Dynamic Prefix Weighting
Hyeonseo Jang, Hyuk Kwon, and Kibok Lee

📖 Overview

DPW (Dynamic Prefix Weighting) is a parameter-efficient framework for continual learning of vision-language models. DPW assigns fine-grained, token-level weights to prefixes and adapters:

• 🔹 A gating module adjusts the weight of each prefix based on the importance of the corresponding input token.
• 🔹 A residual weighting mechanism derives adapter weights from the residuals of the prefix weights.

⚙️ Installation

conda env create -f environment.yml
conda activate dpw

📂 Datasets

DPW is trained and evaluated on the MTIL benchmark from ZSCL. We follow the dataset structure of DIKI; please refer to their dataset instructions for setup details.

🚀 Training

MTIL

Multi-domain Task Incremental Learning, where task IDs are available at test time.

bash MTIL.sh

ODCL-CIL

Open-Domain Continual Learning, where task IDs are not available at test time.

bash ODCL-CIL.sh

Parameter-Efficient Variants

For the reduced-parameter variants (Ours†) reported in the paper:

bash MTIL_reduced_param.sh
bash ODCL-CIL_reduced_param.sh

📝 Citation

If you find this work useful, please consider citing our paper:

@inproceedings{jang2026dpw,
  title={Enhancing Continual Learning of Vision-Language Models via Dynamic Prefix Weighting},
  author={Jang, Hyeonseo and Kwon, Hyuk and Lee, Kibok},
  booktitle={CVPR},
  year={2026}
}

📬 Contact

For questions or issues, please contact jhyeonseo715@yonsei.ac.kr.

🙏 Acknowledgements

This codebase builds on DIKI. We thank the authors for releasing their code.