mlsanityData to 14 Oct 2026

Line of Sight: On Linear Representations in VLLMs

arXiv 2506.04706 ↗First posted 5 Jun 2025cs.CV cs.AIcited by 6

On the map: Mechanistic interpretability → Interpreting vision and multimodal models

Abstract (arXiv)

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.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability. Builds on 22 of its 41 references that are in the map. First posted on arXiv 5 Jun 2025.

Builds on

22 of 41 references.

References · 22

  1. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  2. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  3. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  4. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  6. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  7. 2016The Concrete Distribution: A Continuous Relaxation of Discrete Random Variablescited by 1,148
  8. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  9. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  10. 2017Learning Sparse Neural Networks through L0 Regularizationcited by 583
  11. 2020Understanding the Role of Individual Units in a Deep Neural Networkcited by 255
  12. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
10 more
  1. 2023Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Modelscited by 153
  2. 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Trackingcited by 149
  3. 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 146
  4. 2022Natural Language Descriptions of Deep Visual Featurescited by 137
  5. 2023Mechanistically analyzing the effects of fine-tuning on procedurally defined taskscited by 115
  6. 2018Under the Hood: Using Diagnostic Classifiers to Investigate and Improve how Language Models Track Agreement Informationcited by 97
  7. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
  8. 2021Low-Complexity Probing via Finding Subnetworkscited by 64
  9. 2023Time is Encoded in the Weights of Finetuned Language Modelscited by 33
  10. 2021Linear Semantics in Generative Adversarial Networkscited by 25

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 2; 2026: 4.

6 papers.

  1. 2025Head Pursuit: Probing Attention Specialization in Multimodal Transformerscited by 17
  2. 2025SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucinationcited by 5
  3. 2026Probing Visual Concepts in Lightweight Vision-Language Models for Automated Drivingcited by 3
  4. 2026Language Models Can Explain Visual Features via Steeringcited by 2
  5. 2026Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Modelscited by 1
  6. 2026Visual concept ranking uncovers medical shortcuts used by large multimodal modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders
  2. 2026Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs
  3. 2025Interpreting the linear structure of vision-language model embedding spaces
  4. 2025How Visual Representations Map to Language Feature Space in Multimodal LLMs
  5. 2024The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models
  6. 2024Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features
  7. 2025Explaining How Visual, Textual and Multimodal Encoders Share Concepts
  8. 2025SAE-V: Interpreting Multimodal Models for Enhanced Alignment