Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models (2504.02821v2)

Abstract: Given that interpretability and steerability are crucial to AI safety, Sparse Autoencoders (SAEs) have emerged as a tool to enhance them in LLMs. In this work, we extend the application of SAEs to Vision-LLMs (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in vision 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. Notably, 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 model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised tool for enhancing both interpretability and control of VLMs. Code is available at https://github.com/ExplainableML/sae-for-vlm.