Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Language Models Can Explain Visual Features via Steering

Javier Ferrando, Enrique Lopez-Cuena, Pablo Agustin Martin-Torres, Daniel Hinjos, Anna Arias-Duart, Dario Garcia-Gasulla

Abstract

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Language Models Can Explain Visual Features via Steering

Abstract

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.
Paper Structure (44 sections, 10 equations, 14 figures, 6 tables)

This paper contains 44 sections, 10 equations, 14 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Extracting Features
  3. Automatically Interpreting Features
  4. Top-k Explanations
  5. Proposed Approach
  6. Steering-based Explanations.
  7. Steering-informed Top-k Explanations.
  8. Evaluating the Quality of the Explanations
  9. Evaluation Metrics
  10. Simulation-based Evaluation.
  11. CLIP-based Evaluation.
  12. Synthetic-image-based Evaluation.
  13. Experimental Setup
  14. Results
  15. Explaining through Steering
  16. ...and 29 more sections

Figures (14)

  • Figure 1: Top: A vision feature extracted with an SAE is explained based on top-activating images, which are passed to the VLM to generate an explanation based on correlated visual evidence. Bottom: We propose to automatically obtain explanations of SAE features by causally intervening (steering) a vision encoder. The intervention is done after feeding it an information-devoid white image, effectively making the language model articulate what visual concept that feature represents.
  • Figure 2: Middle layer SAE synthetic-image-based evaluation scores of Top-k method as a function of the similarity with Steering Explanations.
  • Figure 2: Count and percentage of 'background' explanations turned 'animal' explanations by different methods (see main text for details).
  • Figure 3: Gemma 3 synthetic-image-based evaluation scores of Steering method as a function of the size of the LM $\text{m}_{\text{subj}}$.
  • Figure 4: Example of Top-k explanation exhibiting contextual bias.
  • ...and 9 more figures