Sparsity as a Key: Unlocking New Insights from Latent Structures for Out-of-Distribution Detection
Abstract
Sparse Autoencoders (SAEs) have demonstrated significant success in interpreting Large Language Models (LLMs) by decomposing dense representations into sparse, semantic components. However, their potential for analyzing Vision Transformers (ViTs) remains largely under-explored. In this work, we present the first application of SAEs to the ViT [CLS] token for out-of-distribution (OOD) detection, addressing the limitation of existing methods that rely on entangled feature representations. We propose a novel framework utilizing a Top-k SAE to disentangle the dense [CLS] features into a structured latent space. Through this analysis, we reveal that in-distribution (ID) data exhibits consistent, class-specific activation patterns, which we formalize as Class Activation Profiles (CAPs). Our study uncovers a key structural invariant: while ID samples preserve a stable pattern within CAPs, OOD samples systematically disrupt this structure. Leveraging this insight, we introduce a scoring function based on the divergence of core energy profiles to quantify the deviation from ideal activation profiles. Our method achieves strong results on the FPR95 metric, critical for safety-sensitive applications across multiple benchmarks, while also achieving competitive AUROC. Overall, our findings demonstrate that the sparse, disentangled features revealed by SAEs can serve as a powerful, interpretable tool for robust OOD detection in vision models.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Do Vision Transformers See Like Convolutional Neural Networks? cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Swin Transformer: Hierarchical Vision Transformer using Shifted Windows cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Builds on15
- Swin Transformer: Hierarchical Vision Transformer using Shifted WindowsZe Liu, Yutong Lin, Yue Cao, Han Hu et al.ICCV 2021 · 31,683 citations
- An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn et al.ICLR 2021 · 21,477 citations
- Energy-based Out-of-distribution DetectionWeitang Liu, Xiaoyun Wang, John D. Owens, Yixuan LiNeurIPS 2020 · 2,213 citations
- Do Vision Transformers See Like Convolutional Neural Networks?Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang et al.NeurIPS 2021 · 1,553 citations
- Sparse Autoencoders Find Highly Interpretable Features in Language ModelsRobert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart et al.ICLR 2024 · 1,072 citations
Related papers
- Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision TransformersTang Li, Yanlin Chen, Mengmeng Ma, Xi PengICML 2026
- Sparse Autoencoders Trained on the Same Data Learn Different FeaturesGonçalo Paulo, Nora BelroseICLR 2026 · 96 citations
- DAVE: Distribution-aware Attribution via ViT Gradient DecompositionAdam Wróbel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele et al.ICML 2026 · 2 citations
- A is for Absorption: Studying Feature Splitting and Absorption in Sparse AutoencodersDavid Chanin, James Wilken-Smith, Tomás Dulka, Hardik Bhatnagar et al.NeurIPS 2025 · 168 citations
- TIDE: Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image GenerationVictor Shea-Jay Huang, Le Zhuo, Yi Xin, Zhaokai Wang et al.AAAI 2026 · 10 citations
