Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIP

Yusung Ro; Jaehyun Choi; Junmo Kim

cs.CVApr 7, 2026

Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIP

Authors: Yusung Ro, Jaehyun Choi, Junmo Kim

Organizations: KAIST · Korea AI Safety Institute, ETRI

Abstract

Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpreting the internal representations of CLIP vision encoders, yet existing analyses largely focus on the semantic meaning of individual features. We introduce information scope as a complementary dimension of interpretability that characterizes how broadly an SAE feature aggregates visual evidence, ranging from localized, patch-specific cues to global, image-level signals. We observe that some SAE features respond consistently across spatial perturbations, while others shift unpredictably with minor input changes, indicating a fundamental distinction in their underlying scope. To quantify this, we propose the Contextual Dependency Score (CDS), which separates positionally stable local scope features from positionally variant global scope features. Our experiments show that features of different information scopes exert systematically different influences on CLIP's predictions and confidence. These findings establish information scope as a critical new axis for understanding CLIP representations and provide a deeper diagnostic view of SAE-derived features.

Figures & tables

Figure 1: Spatial instability of CLIP outlier tokens under Shifted Context Cropping (SCC). We analyze the spatial consistency of patch-token L2 norms using two overlapping crops (red: top-left; blue: bottom-right) translated by a single patch. The heatmaps visualize patch-token L2 norms for each crop across different CLIP ViT models. The green and yellow spots, indicating outlier tokens, do not maintain consistent spatial locations between the two overlapping crops despite this minimal contextual shift. This shows that outlier token locations are highly unstable under SCC.

Figure 2: Framework for quantifying our Contextual Dependency Score (CDS) of SAE features. We evaluate the contextual dependency of SAE features using CDS. This score is derived from the Earth Mover’s Distance (EMD) computed between the normalized activation maps of two shifted input crops (red and blue) as shown for a single pair. The final CDS for a feature is the average of these EMD scores over $k_{CDS}$ representative images. We process only the embeddings from the overlapping region.

Panel of Figure 3 — Figure 3: Distribution of Contextual Dependency Score (CDS). Histogram of CDS values for SAE features across three CLIP models, showing a dominant low-CDS peak together with distinct peaks in the high-CDS region. This distribution motivates partitioning SAE features into low-CDS and high-CDS groups.

Model	Head	EMD
Model	Head	non-outlier	outlier
CLIP-B/16	H11 ( $D_{11}$ )	$1.365_{\pm.28}$	$5.415_{\pm 1.52}$
CLIP-B/16	All heads ( $\bar{D}$ )	$2.495_{\pm.59}$	$5.389_{\pm 1.54}$
CLIP-L/14	H4 ( $D_{4}$ )	$0.796_{\pm.41}$	$4.398_{\pm 3.33}$
CLIP-L/14	All heads ( $\bar{D}$ )	$1.329_{\pm 1.20}$	$4.411_{\pm 3.35}$
CLIP-L/14-336px	H4 ( $D_{4}$ )	$0.903_{\pm.47}$	$5.164_{\pm 4.21}$

Table 1: Quantitative analysis of attention instability for outlier vs. non-outlier tokens. We report the EMD scores (

\bar{D}_{\text{out}}

vs.

\bar{D}_{\text{non}}

) computed on the final layer’s attention maps using our SCC method with a minimal shift

s=1

to induce perturbation. The comparison is provided for three CLIP models, averaging across all heads and isolating semantic heads.

Model	Feature Set	Non-Outlier Token	Outlier Token
CLIP-B/16	high-CDS	$1.66_{\pm.97}$	$83.45_{\pm 3.19}$
CLIP-B/16	low-CDS	$31.51_{\pm 4.02}$	$10.39_{\pm 3.22}$
CLIP-L/14	high-CDS	$2.69_{\pm.56}$	$171.55_{\pm 16.47}$
CLIP-L/14	low-CDS	$52.44_{\pm 5.77}$	$25.40_{\pm 5.66}$
CLIP-L/14-336px	high-CDS	$8.94_{\pm 1.50}$	$173.65_{\pm 16.22}$
CLIP-L/14-336px	low-CDS	$50.89_{\pm 5.11}$	$89.31_{\pm 8.64}$

Table 2: Feature set activations by token type. We compare the average activation of the low-CDS set and high-CDS set in response to outlier tokens versus non-outlier tokens.

Model	Embedding Type	ImageNet1K Top-1↑	ADE20K mIoU↑	NYUd rmse↓
CLIP-B/16	Original high-CDS-removed low-CDS-removed	74.82 75.54 64.86	25.87 26.02 11.65	0.8841 0.8616 0.9481
CLIP-L/14	Original high-CDS-removed low-CDS-removed	80.82 81.28 78.30	26.66 26.44 13.89	0.8029 0.7994 0.8878
CLIP-L/14-336px	Original high-CDS-removed low-CDS-removed	80.69 82.11 82.89	24.89 24.00 15.48	0.8063 0.7894 0.8825

Table 3: Linear probing performance of ablated feature sets. We report linear probing results (ImageNet1K Top-1, ADE20K mIoU, NYUd RMSE) across different CLIP backbones. We compare the performance of the original embedding against two ablations: removing low-CDS features and removing high-CDS features. BOLD denote the best result, while UNDERLINE denote the worst result.

Figure 4: Representative activation patterns of low-CDS and high-CDS SAE features. Low-CDS features (left) show spatially grounded and semantically consistent responses, whereas high-CDS features (right) exhibit more diffuse activations and weaker semantic consistency.

Figure 5: Visualization of per-patch $L_{2}$ norms for the original and feature-removed embeddings. We visualize the per-patch $L_{2}$ norms of the original embedding, the low-CDS-removed embedding, and the high-CDS-removed embedding. The comparison highlights how each feature group affects outlier-token norms and the overall spatial distribution of patch norms across the three CLIP models.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Dictionary Size	$\gamma$	Low-CDS	High-CDS
CLIP-B/16	24,576	0.14	17,329	7,247
CLIP-L/14	32,768	0.20	27,887	4,881
CLIP-L/14-336px	32,768	0.13	22,323	10,445

Appendix

Table 4: SAE dictionary sizes and CDS-based partition statistics for the CLIP vision encoders used in the main paper. The high-CDS set forms a relatively small subset of the full SAE dictionary across all three CLIP backbones.

Architecture	Model	Layer Used	$\tau$
	Base	Penultimate (11th)	100
DINOv2	Large	Penultimate (23rd)	100
	Giant	Last (40th)	150
	Small	10th	500
DeiT3	Base	9th	500
	Large	18th	1000

Appendix

Table 5: Model-specific SAE configurations for the additional ViT architectures. We report the transformer layer used for SAE training and analysis, together with the empirically selected outlier threshold

\tau

for each model.

Model	FVU A $\downarrow$	FVU O $\downarrow$	$L_{0}{}_{A}$	$L_{0}{}_{O}$	CS A $\uparrow$	CS O $\uparrow$
CLIP-B/16	.099	.021	30.201	13.064	.923	.998
CLIP-L/14	.176	.042	29.307	14.170	.873	.999
CLIP-L/14-336px	.115	.028	29.654	16.391	.913	.999
DINOv2-B	.039	.001	29.425	7.298	.911	.999
DINOv2-L	.115	.001	30.416	15.334	.917	.999
DINOv2-G	.062	.002	30.664	28.709	.912	.999

Appendix

Table 6: SAE reconstruction statistics for all tokens (A) and outlier tokens (O).

Panel of Figure 6 — Figure 6: Distribution of Contextual Dependency Score (CDS). Histogram of CDS values for SAE features across DINOv2, SigLIP2, and DeiT3.

Model	non-outlier ( $\bar{D}_{\text{non}}$ )	outlier ( $\bar{D}_{\text{out}}$ )
DINOv2	0.61	6.96
DeiT3	0.38	5.97
SigLIP2	3.75	7.18

Appendix

Table 7: Contextual instability of attention maps for outlier vs. non-outlier tokens. Average EMD scores are computed on the designated layer’s attention maps using our SCC method (

s=1

). Across all models, outlier tokens exhibit significantly higher EMD scores, demonstrating severe sensitivity to contextual shifts.

Figure 7: Contextual dependency across the norm spectrum. We plot the Activation-Weighted CDS ( awCDS ) across the token norm spectrum. Across all architectures, awCDS spikes exponentially for outlier tokens.

Figure 8: Linear probe accuracy across ViT scales on ImageNet-1K. The high-CDS-removed embedding (blue) consistently outperforms the Baseline (green) across all architectures, indicating that local signals are more discriminative for classification. Conversely, the performance gap of the low-CDS-removed embedding (red) generally narrows as model size increases, suggesting that larger models encode broader context more effectively.

Model	Embedding Type	ADE20K mIoU↑	NYUd rmse↓
DINOv2	Original high-CDS-removed low-CDS-removed	25.26 26.11 15.47	0.7500 0.7539 0.8402
DeiT3	Original high-CDS-removed low-CDS-removed	26.48 26.26 8.38	0.8444 0.8391 1.0607
SigLIP2	Original high-CDS-removed low-CDS-removed	35.62 35.39 16.93	0.8178 0.8123 0.9205

Appendix

Table 8: Linear probing performance of ablated feature sets on other ViT architectures. We report linear probing results (ADE20K mIoU, NYUd RMSE) for the Base-scale DINOv2, DeiT3, and SigLIP2 models. We compare the original embedding against two ablations: removing low-CDS features and removing high-CDS features. BOLD denotes the best result, while UNDERLINE denotes the worst result.

Shifting factor $s$	1	2	3	4	5	6
Non-outlier	1.490	1.745	1.842	1.855	1.915	1.958
Outlier	5.179	6.752	6.869	6.876	5.811	5.533

Appendix

Table 9: Sensitivity of non-outlier and outlier tokens under varying shifting factors in SCC. Outlier tokens consistently exhibit higher instability than non-outlier tokens across all values of

s

, the shifting factor in Shifted Context Crop (SCC).

Figure 9: Visualization of low-CDS features. These features exhibit strong spatial grounding, consistently localizing specific visual concepts across different images.

Figure 10: Visualization of high-CDS features. These features exhibit diffuse activation patterns and capture broader contextual information rather than localized visual details.

Neural Network Interpretability

Neural network interpretability studies how models represent information and make predictions, and how to explain, test, or control their internal computations.

See all 330 papers

ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection
Oct 4, 2026Qingwen Zeng, Zehao Fu, Shuyu Meng +5Transformer Interpretability Sparse Autoencoders
The University of Sydney · The Hong Kong University of Science and Technology (Guangzhou) · DeepWisdom +1
Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects
Jul 22, 2026Seonglae Cho, Zekun Wu, Kleyton Da Costa +3LLM Interpretability Sparse Autoencoders
Holistic AI · University College London
Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects
Jul 27, 2026Phu Gia Hoang, Anwoy Chatterjee, Tanmoy Chakraborty +2Sparse Autoencoders Mechanistic Interpretability
UKP Lab, Technical University of Darmstadt · Department of Electrical Engineering Indian Institute of Technology Delhi, India

Oct 4, 2026J/K move · Enter open · S save

ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection

Qingwen Zeng, Zehao Fu, Shuyu Meng, Linghan Huang+4

The University of Sydney · The Hong Kong University of Science and Technology (Guangzhou) · DeepWisdom+1

Transformer Interpretability Sparse Autoencoders Mechanistic Interpretability

Open paper AskPDF arXiv

Sparse autoencoders (SAEs) are a central tool in mechanistic interpretability. However, existing SAEs are primarily trained per layer. The modeling subspace is therefore fixed by layer identity, independent of which token-layer states actually drive each prediction. We argue that this constraint contributes to several limitations observed in layer-wise SAEs, including low feature utilization, high dictionary redundancy, and features that lack direct behavioral grounding. In this paper, we propose ScopeSAE, which selects the modeling subspace per token by attributing each prediction to its most influential token-layer state via normalized gradient-based attribution, and learns features over the resulting prediction-relevant subspace. Empirically, ScopeSAE yields an effect we term reconstruction-better-than-original. Written-back reconstructions of the SAE produce lower next-token cross-entropy than the original activations, an outcome that, to our knowledge, has not previously been reported for SAEs. Through interventional analyses and a KL fine-tuning counter-experiment, we show that this effect is attributable to ScopeSAE's prediction-relevant subspace itself rather than to architectural changes. ScopeSAE further improves effective feature count, interpretability, utilization, and dictionary redundancy over existing layer-based baselines, suggesting that choosing the SAE modeling subspace by predictive relevance leads to more useful and behaviorally meaningful features.

Abstract

Neural Network Interpretability

Explore similar work

ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection

Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects

Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects