Line of Sight: On Linear Representations in VLLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Rajaram, Achyuta, Schwettmann, Sarah, Andreas, Jacob, Conmy, Arthur
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910989248626688
author Rajaram, Achyuta
Schwettmann, Sarah
Andreas, Jacob
Conmy, Arthur
author_facet Rajaram, Achyuta
Schwettmann, Sarah
Andreas, Jacob
Conmy, Arthur
contents Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show that the features are causal by performing targeted edits on the model output. In order to increase the diversity of the studied linear features, we train multimodal Sparse Autoencoders (SAEs), creating a highly interpretable dictionary of text and image features. We find that although model representations across modalities are quite disjoint, they become increasingly shared in deeper layers.
format Preprint
id arxiv_https___arxiv_org_abs_2506_04706
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Line of Sight: On Linear Representations in VLLMs
Rajaram, Achyuta
Schwettmann, Sarah
Andreas, Jacob
Conmy, Arthur
Computer Vision and Pattern Recognition
Artificial Intelligence
Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show that the features are causal by performing targeted edits on the model output. In order to increase the diversity of the studied linear features, we train multimodal Sparse Autoencoders (SAEs), creating a highly interpretable dictionary of text and image features. We find that although model representations across modalities are quite disjoint, they become increasingly shared in deeper layers.
title Line of Sight: On Linear Representations in VLLMs
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2506.04706