🚀 $A_3B_2$: Adaptive Asymmetric Adapter for Alleviating Branch Bias in Vision-Language Image Classification with Few-Shot Learning (IJCAI2026)

PyTorch implementation of $A_3B_2$

🔧Installation

1. Install the dassl environment by following the instructions on Dassl.pytorch and install PyTorch. Then, run pip install -r requirements.txt to install additional packages for CLIP (make sure dassl is activated).
2. Follow DATASETS.md to install the datasets.

✨Experiments

• 😃Motivation (Branch Bias Experiments)

  • Base-to-Novel Generalization

    bash motivation_b2n.sh

  • Cross-Dataset Evaluation & Domain Generalization

    bash motivation_cross.sh

• 🥳Main Experiments

  • Base-to-Novel Generalization (Few-Shot Learning)

    bash base_to_novel.sh

  • Cross-Dataset Evaluation & Domain Generalization

    bash cross_datasets.sh

🥹Citation

If you find our work valuable, we would greatly appreciate your citation:

@misc{zhou2026a3b2adaptiveasymmetricadapter,
      title={A$_3$B$_2$: Adaptive Asymmetric Adapter for Alleviating Branch Bias in Vision-Language Image Classification with Few-Shot Learning}, 
      author={Yiyun Zhou and Zhonghua Jiang and Wenkang Han and Kunxi Li and Mingjing Xu and Chang Yao and Jingyuan Chen},
      year={2026},
      eprint={2605.13161},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.13161}, 
}