Language Models Can Explain Visual Features via Steering

🕒 Published (v1): 2026-03-23 21:37 UTC · Source: Arxiv · Venue: CVPR 2026 · link

Why this paper was selected

CVPR 2026; LLMs explain SAE visual features autonomously via causal steering

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper introduces Steering, a causal intervention method that explains Sparse Autoencoder (SAE) features in vision encoders by injecting individual SAE decoder vectors into the residual stream of a blank-image forward pass and prompting the attached language model to describe what it "sees." A hybrid variant, Steering-informed Top-k, combines this causal signal with top-activating images to achieve state-of-the-art explanation quality at no additional computational cost.

Problem

Existing automated interpretability methods for vision model SAE features rely on correlation-based explanations derived from top-activating images, which (1) require a large labeled evaluation set, (2) introduce contextual bias (e.g., misattributing background features to foreground animals due to recurring co-occurring objects), and (3) cannot causally ground the explanations in feature identity. Thousands of SAE features in scaled vision models remain practically uninterpretable without human intervention.

Method

The method exploits the architecture of VLMs (frozen vision encoder + language model backbone):

1. Steering: Feed an information-devoid white image \(\tilde{I}\) into the vision encoder. At layer \(l\), replace the residual stream across all patch positions via the do-operator: $\(e_i \sim m_{\text{exp}}\!\left(e \mid P,\, \tilde{I},\, \mathrm{do}\!\left(m^l_{\text{sub}}(\tilde{I}) \leftarrow m^l_{\text{sub}}(\tilde{I}) + \alpha W_{\text{dec}}[i, :]\right)\right)\)$ where \(W_{\text{dec}}[i,:]\) is the \(i\)-th row of the SAE decoder matrix and \(\alpha\) is a validation-tuned coefficient. The LM then generates a natural-language description of the injected feature. Requires a single forward pass; no image dataset needed.

2. Steering-informed Top-k: Apply the same causal intervention while conditioning on the top-\(k\) activating images \(T^k_i\): $\(e_i \sim m_{\text{exp}}\!\left(e \mid P,\, T^k_i,\, \mathrm{do}\!\left(m^l_{\text{sub}}(T^k_i) \leftarrow m^l_{\text{sub}}(T^k_i) + \alpha W_{\text{dec}}[i, :]\right)\right)\)$

SAEs used are TopK SAEs (\(d_{\text{SAE}} = 8{,}192\)) trained on ImageNet, applied to middle layers of Gemma 3 (SigLIP 400M encoder) and InternVL3-14B (InternViT-300M encoder).

Evaluation uses four metrics: IoU-Score (SAM-2 segmentation vs. actual activation masks), CLIP-Score (cosine similarity of explanation text embedding to masked image embeddings), Synthetic Activation Score (Stable Diffusion–generated images fed back to measure feature activation), and AUROC (positive vs. randomly sampled negative images).

Key Contributions

• Causal steering for visual feature explanation: first method to causally intervene on SAE features in a VLM vision encoder and use the LM head as an automated explainer, requiring no evaluation image set.
• Steering-informed Top-k: hybrid combining causal intervention with top-activating images, achieving best performance across all four metrics and both encoder families without extra compute.
• Contextual bias analysis: quantifies how Top-k misattributes background features (up to 47.7% "animal" labels for heatmap-based Top-k) and shows Steering reduces this.
• Scaling law for explanation quality: demonstration that Steering quality improves monotonically with LM size (4B → 12B → 27B Gemma 3).
• Semantic filtering via Steering similarity: Top-k explanations with high semantic similarity to Steering explanations are consistently higher quality (validated on both AUROC and Synthetic Activation Score).
• SAE feature space visualization: UMAP/k-means analysis of ~5,000 features revealing low-level perceptual and high-level semantic clusters; discovery of depth/perspective features in SigLIP previously thought unique to DINOv2.

Results

• Steering-informed Top-k outperforms both baselines on all metrics (Gemma 3 middle layer, Masks): IoU 0.216 vs. 0.211 (Top-k) vs. 0.220 (Steering); AUROC 0.788 vs. 0.723 vs. 0.655; Mean Synthetic Activation 0.461 vs. 0.330 vs. 0.141; CLIP Score 0.193 vs. 0.190 vs. 0.191.
• InternVL3-14B results mirror the pattern: Steering-informed Top-k achieves AUROC 0.823 vs. 0.768 (Top-k); Synthetic Activation 0.254 vs. 0.187.
• Steering scales with LM size: AUROC and Mean Synthetic Activation increase monotonically from 4B to 12B to 27B Gemma 3 LM.
• Contextual bias: Top-k (heatmaps) labels 47.7% of high-IoU-Steering features as "animal"; Steering itself labels 0%; Steering-informed Top-k reduces this to 38.6%.
• Steering-informed Top-k overcomes diminishing returns from dataset scaling: achieves high quality even at 1k images, surpassing standard Top-k using 50k images on synthetic metrics.
• ~5,000 of 7,690 SAE features pass quality thresholds; Steering-informed Top-k is selected most frequently by rank-based voting.

Limitations

• Steering underperforms Top-k on AUROC and Synthetic Activation Score when used in isolation, indicating the blank-image causal signal alone is insufficient for full explanation fidelity.
• Evaluation is conducted on ImageNet-trained SAEs; generalization to other domains or architectures is untested.
• The steering coefficient \(\alpha\) is selected on a 500-feature validation subset, which may not transfer across significantly different feature distributions.
• Causal interpretation assumes the VLM language model reliably surfaces the injected visual concept; confounds from LM priors are not fully characterized.
• Claims about depth/perspective features unique to DINOv2 are noted as anecdotal; no systematic cross-model comparison is performed.

Relevance to Vision-Language Models

This work directly targets the interpretability of the vision encoder component in VLMs, using the language model head as an oracle to decode what individual SAE features encode—making internal representations of VLMs more transparent without human labeling. The finding that Steering quality scales with LM size implies that as VLMs grow, their own generative capacity becomes an increasingly powerful auto-interpretability tool, forming a self-reinforcing loop between model capability and mechanistic transparency. For researchers tracking VLMs, the contextual bias finding is practically significant: it suggests that standard feature-attribution pipelines built on top-activating image sets may systematically mischaracterize what VLM visual encoders actually represent. The hybrid Steering-informed Top-k provides a drop-in improvement to any existing VLM interpretability pipeline.