Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models

Mateusz Pach, Shyamgopal Karthik, Quentin Bouniot, Serge Belongie, Zeynep Akata

Abstract

Sparse Autoencoders (SAEs) have recently gained attention as a means to improve the interpretability and steerability of Large Language Models (LLMs), both of which are essential for AI safety. In this work, we extend the application of SAEs to Vision-Language Models (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in visual representations. To ensure that our evaluation aligns with human perception, we propose a benchmark derived from a large-scale user study. Our experimental results reveal that SAEs trained on VLMs significantly enhance the monosemanticity of individual neurons, with sparsity and wide latents being the most influential factors. Further, we demonstrate that applying SAE interventions on CLIP's vision encoder directly steers multimodal LLM outputs (e.g., LLaVA), without any modifications to the underlying language model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised tool for enhancing both interpretability and control of VLMs.

---

Setup

Install required PIP packages.

pip install -r requirements.txt

Download following datasets:

• ImageNet (https://pytorch.org/vision/main/generated/torchvision.datasets.ImageNet.html)
• INaturalist 2021 (https://github.com/visipedia/inat_comp/tree/master/2021)

Export paths to dataset directories. The directories should contain train/ and val/ subdirectories.

export IMAGENET_PATH="<path_to_imagenet>"
export INAT_PATH="<path_to_inaturalist>"

Code was run using Python version 3.11.10.

Running Experiments

The commands required to reproduce the results are organized into scripts located in the scripts/ directory:

• monosemanticity_score.sh computes the Monosemanticity Score (MS) for specified SAEs, layers, models, and image encoders.
• matryoshka_hierarchy.sh analyzes the hierarchical structure that emerges in Matryoshka SAEs.
• mllm_steering.sh enables experimentation with steering LLaVA using an SAE built on top of the vision encoder.

Metric Benchmark

To facilitate advancements in monosemanticity metrics, we release benchmark data derived from our user study in the metric_benchmark.tar.gz archive, comprising:

• pairs.csv – Contains 1000 pairs of neurons (r_x, r_y), along with user preferences R_user and MS values computed using two different image encoders: DINOv2 ViT-B and CLIP ViT-B. Each row includes the following columns:
  k_x, k_y, R_user, MS_x_dino, MS_y_dino, MS_x_clip, MS_y_clip.

• top16_images.csv – Lists the 16 most activating images from the ImageNet training set for each neuron used in the study. Columns:
  k, x_1, …, x_16.

• activations.csv – Provides activation values of all 50,000 ImageNet validation images for each neuron. Columns:
  k, a_1, …, a_50000.

See Appendix D: Benchmark attached to the article to learn more.

We use the implementation of sparse autoencoders available at https://github.com/saprmarks/dictionary_learning.

Citation

@article{pach2025sparse,
  title={Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models}, 
  author={Mateusz Pach and Shyamgopal Karthik and Quentin Bouniot and Serge Belongie and Zeynep Akata},
  journal={arXiv preprint arXiv:2504.02821},
  year={2025}
}