mateuszpach/sae-for-vlm

Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models

This repository contains the checkpoints of sparse autoencoders trained in the paper Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models.

Code: https://github.com/ExplainableML/sae-for-vlm

How to download the weights:

from huggingface_hub import login, snapshot_download

login()

snapshot_download(
    repo_id="mateuszpach/sae-for-vlm",
    repo_type="model",
    local_dir="checkpoints_dir",
)

Citation

@article{pach2026sparse,
  title={Sparse autoencoders learn monosemantic features in vision-language models},
  author={Pach, Mateusz and Karthik, Shyamgopal and Bouniot, Quentin and Belongie, Serge and Akata, Zeynep},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  pages={95706--95742},
  year={2026}
}