Language Models Can Explain Visual Features via Steering¶

Conference: CVPR 2026 arXiv: 2603.22593 Code: GitHub Area: Interpretability Keywords: Sparse Autoencoders, Visual Feature Explanation, Causal Intervention, VLM, Automated Interpretability

TL;DR¶

This paper proposes a method for scalable automatic explanation of visual features by causally intervening (steering) on SAE features in VLM visual encoders. By injecting feature vectors into a blank image's forward pass and prompting the language model to describe what it "sees," the approach eliminates the need for an evaluation image set. A hybrid method, Steering-informed Top-k, is further proposed and achieves state-of-the-art performance.

Background & Motivation¶

Root Cause¶

Key Challenge: Background: Sparse autoencoders (SAEs) have emerged as a powerful tool for discovering interpretable features in vision models. However, automatically explaining thousands of discovered features at scale remains an open problem.

Limitations of existing methods (Top-k approaches): 1. Correlation-based rather than causal: Selecting highest-activating images and identifying common patterns is fundamentally a correlational analysis. 2. Dependence on evaluation image sets: Retrieving top-activating images requires large-scale datasets, introducing dataset bias. 3. High computational cost: Ranking activations requires a full forward pass over the entire evaluation set.

Core insight of this paper: VLMs connect a visual encoder to a pretrained language model. If a causal intervention is applied to the visual encoder — injecting a specific SAE feature vector into a blank image's forward pass — the language model should be able to express what visual concept it "perceives."

Method¶

Overall Architecture¶

Train a TopK SAE (\(d_{\text{SAE}} = 8192\)) to decode visual encoder features on ImageNet.
For each SAE feature, inject the corresponding feature vector during the forward pass of a blank image.
Prompt the language model to explain what it "sees."
Optionally: apply a hybrid approach that combines Top-k images with causal intervention.

Key Designs¶

Steering-based Explanation (pure causal intervention):
- Function: Generates natural language explanations of SAE features without requiring an image set.
- Mechanism: A blank (all-white) image \(\tilde{I}\) is fed into the VLM. At layer \(l\) of the visual encoder, the SAE decoder weight vector \(W_{dec}[i,:] \times \alpha\) is added to the residual stream at all token positions. The language model then generates an explanation conditioned on the intervened visual representation, formalized as: \(e_i \sim m_{exp}(e | P, \tilde{I}, \text{do}(m_{sub}^l(\tilde{I}) \leftarrow m_{sub}^l(\tilde{I}) + \alpha W_{dec}[i,:]))\)
- Design Motivation: A blank image provides no meaningful visual signal, so the language model's output is entirely driven by the causal intervention, yielding a purely causal explanation. Only a single forward pass is required, making the approach highly efficient.
Steering-informed Top-k (hybrid method):
- Function: Combines the advantages of causal intervention and input images.
- Mechanism: The same SAE feature causal intervention is applied to the visual encoder while also conditioning on Top-k activating images. This integrates correlational evidence (Top-k images) with causal evidence (feature injection) to guide the explainer toward more precise descriptions.
- Design Motivation: Pure steering performs better on low-level features, while Top-k is stronger for high-level semantic features. The hybrid method achieves optimal performance across all four complementary metrics without incurring additional computational cost.
Evaluation Metric Suite:
- Function: Quantifies explanation quality from multiple complementary perspectives.
- Four complementary metrics are adopted:
  - Activation IoU: Overlap between images highly activated by the explanation text and those highly activated by the SAE feature.
  - Detection Score: Whether the described concept can be detected by a VLM in the activating images.
  - CLIP Similarity: CLIP embedding distance between the explanation and the top-activating images.
  - Monosemanticity: Whether the feature corresponds to a single coherent concept.

Loss & Training¶

The SAE is trained with a standard TopK objective on ImageNet. The intervention strength \(\alpha\) is selected on a validation set of 500 features. The explanation generation procedure itself involves no training — it operates purely at inference time via intervention.

Key Experimental Results¶

Main Results — Explanation Quality Comparison (Gemma 3 Visual Encoder)¶

Method	Activation IoU↑	Detection↑	CLIP↑	Monosemanticity↑
Top-k (original images)	baseline	baseline	baseline	baseline
Top-k (Mask)	slightly better	slightly better	slightly lower	comparable
Top-k (Heatmap)	comparable	comparable	comparable	comparable
Steering (pure intervention)	below Top-k	below Top-k	below Top-k	comparable
Steering-informed Top-k	best	best	best	best

Ablation Study — Effect of Language Model Scale¶

LM Scale	Explanation Quality Trend
Small	baseline level
Medium	significant improvement
Large	continued improvement

Explanation quality improves consistently with language model scale, with no sign of saturation.

Key Findings¶

Pure steering outperforms Top-k on low-level features (texture, color, edges), as causal intervention more effectively captures primitive visual concepts.
Top-k is stronger on high-level semantic features (object categories), as concrete images provide useful reference context.
The hybrid method (Steering-informed Top-k) achieves state-of-the-art across all metrics with no additional computational overhead.
Language model scale is a key factor for explanation quality — larger LMs more effectively "verbalize" visual concepts.
Findings are consistent across two distinct VLMs: Gemma 3 and InternVL3.

Highlights & Insights¶

A paradigm shift from correlation to causation: Steering directly intervenes on internal model representations, providing a more causally grounded basis than the correlational analysis of Top-k methods.
Highly efficient: explaining a single feature requires only one forward pass, with no need to iterate over an evaluation set.
The scaling effect of language models suggests that stronger future LMs will further advance automated interpretability.
The hybrid method is elegantly designed: causal signals are injected within the image context provided by Top-k, allowing the two sources of evidence to complement each other.

Limitations & Future Work¶

Pure steering underperforms Top-k on high-level semantic features due to the lack of contextual information in blank images.
The intervention strength \(\alpha\) is sensitive and requires tuning on a validation set.
Validation is limited to VLM architectures; the approach cannot be directly applied to pure vision models without a language model component.
The SAE dictionary size is fixed at 8192; the scalability to larger dictionaries remains unknown.
Evaluation relies primarily on automated metrics, with no human evaluation.

vs. standard Top-k methods: Top-k is correlation-based, requires an evaluation set, and is computationally intensive; Steering is causal, requires no image set, and needs only a single forward pass.
vs. PatchScopes/SELFIE: These methods perform self-explanation within language models; this paper is the first to extend the paradigm to visual encoders.
vs. CB-SAE (same venue): CB-SAE focuses on controllability and interpretability metrics of SAEs, whereas this paper focuses on generating natural language explanations for SAE features.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ — Causal intervention for visual feature explanation is a genuinely novel paradigm; the method is concise and elegant.
Experimental Thoroughness: ⭐⭐⭐⭐ — Multi-metric, multi-VLM, and scaling analysis are provided, though human evaluation is absent.
Writing Quality: ⭐⭐⭐⭐ — Motivation is clear and the method is intuitive, though some LaTeX rendering issues are present.
Value: ⭐⭐⭐⭐ — Makes a significant contribution to automated interpretability research for vision models, with strong scalability.