When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities

• URL: http://arxiv.org/abs/2607.08605v1
• Date: Thu, 09 Jul 2026 15:35:08 GMT
• Title: When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities
• Abstract summary: We propose a Structured Sparse AutoEncoder ($S2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality.<n>We group image patches based on Transformer attention similarity and spatial proximity, and introduce a structured sparsity regularization when training the vanilla SAE.<n>The regularization consists of exclusive sparsity for inter-group concept disentanglement and group sparsity for intra-group concept consistency, which drives the latent neurons by SAEs to specialize in distinct, semantically grounded concepts.
• Score: 10.807285909852277
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept. However, in vision-language models (VLMs), vanilla SAEs struggle to learn modality-consistent concepts, with concepts often exhibiting fragmented coverage (i.e., disjoint regions) in the visual modality. To address this challenge, we propose a Structured Sparse AutoEncoder ($S^2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality. Specifically, we group image patches based on Transformer attention similarity and spatial proximity, and introduce a structured sparsity regularization when training the vanilla SAE. The regularization consists of exclusive sparsity for inter-group concept disentanglement and group sparsity for intra-group concept consistency, which drives the latent neurons by SAEs to specialize in distinct, semantically grounded concepts. Evaluated on the \texttt{Qwen2.5-VL-7B-Instruct} model, the method achieves 6.06% average improvement in semantic alignment (mIoU) and 60.81 in representational efficiency (lower l0 norm) while maintaining near-perfect reconstruction fidelity with an Explained Variance above 99%. Cross-modal analysis further demonstrates that $S^2AE$ enhances neuronal monosemanticity by this visual structural prior, achieving a 3.08% average gain in semantic consistency and a 2.37% average gain in monosemanticity scores for both modalities of multimodal features, thereby fostering more coherent and disentangled representations.

Related papers

• Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders [11.191047696392026]
  Vision-language models map images and text into a joint embedding space.<n>These embeddings often entangle multiple semantic features, which limits their interpretability and controllability.<n>We propose an approach that trains modality-specific sparse autoencoders to preserve each modality's feature geometry, and then aligns corresponding features post hoc.
  arXiv  Detail & Related papers  (2026-06-29T07:27:24Z)
• Towards Foundation Models for 3D Scene Understanding: Instance-Aware Self-Supervised Learning for Point Clouds [53.82500407523346]
  PointINS is an instance-oriented self-supervised framework that enriches point cloud representations through geometry-aware learning.<n>PointINS achieves on average +3.5% mAP improvement for indoor instance segmentation and +4.1% PQ gain for outdoor panoptic segmentation.
  arXiv  Detail & Related papers  (2026-03-26T08:31:06Z)
• SCoCCA: Multi-modal Sparse Concept Decomposition via Canonical Correlation Analysis [7.4746628878102195]
  We introduce Concept CCA (CoCCA), a framework that aligns cross-modal embeddings while enabling interpretable concept decomposition.<n>We propose Sparse Concept CCA (SCoCCA), which enforces sparsity to produce more disentangled and discriminative concepts.<n>Our approach generalizes concept-based explanations to multi-modal embeddings and facilitates state-of-the-art performance in concept discovery.
  arXiv  Detail & Related papers  (2026-03-14T10:45:59Z)
• Concept-Guided Fine-Tuning: Steering ViTs away from Spurious Correlations to Improve Robustness [16.541207762213272]
  Vision Transformers (ViTs) often degrade under distribution shifts because they rely on spurious correlations.<n>We introduce a novel finetuning framework that steers model reasoning toward concept-level semantics.<n>Our method improves robustness across multiple ViT-based models.
  arXiv  Detail & Related papers  (2026-03-09T12:31:14Z)
• Perception-Aware Multimodal Spatial Reasoning from Monocular Images [57.42071289037214]
  spatial reasoning from monocular images is essential for autonomous driving.<n>Current Vision-Language Models (VLMs) still struggle with fine-grained geometric perception.<n>We propose a perception-aware multimodal reasoning framework that equipsVLMs with explicit object-centric grounding ability.
  arXiv  Detail & Related papers  (2026-03-07T02:05:12Z)
• SSVP: Synergistic Semantic-Visual Prompting for Industrial Zero-Shot Anomaly Detection [55.54007781679915]
  We propose Synergistic Semantic-Visual Prompting (SSVP), that efficiently fuses diverse visual encodings to elevate model's fine-grained perception.<n>SSVP achieves state-of-the-art performance with 93.0% Image-AUROC and 92.2% Pixel-AUROC on MVTec-AD, significantly outperforming existing zero-shot approaches.
  arXiv  Detail & Related papers  (2026-01-14T04:42:19Z)
• VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set [80.50996301430108]
  The alignment of vision-language representations endows current Vision-Language Models with strong multi-modal reasoning capabilities.<n>We propose VL-SAE, a sparse autoencoder that encodes vision-language representations into its hidden activations.<n>For interpretation, the alignment between vision and language representations can be understood by comparing their semantics with concepts.
  arXiv  Detail & Related papers  (2025-10-24T10:29:31Z)
• Analyzing Latent Concepts in Code Language Models [10.214183897113118]
  We propose Code Concept Analysis (CoCoA): a global post-hoc interpretability framework.<n>CoCoA uncovers emergent lexical, syntactic, and semantic structures in a code language model's representation space.<n>We propose a hybrid annotation pipeline that combines static analysis tool-based syntactic alignment with prompt-engineered large language models.
  arXiv  Detail & Related papers  (2025-10-01T03:53:21Z)
• Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts [79.18608192761512]
  Self-Explainable Models (SEMs) rely on Prototypical Concept Learning (PCL) to enable their visual recognition processes more interpretable.<n>We propose a Few-Shot Prototypical Concept Classification framework that mitigates two key challenges under low-data regimes: parametric imbalance and representation misalignment.<n>Our approach consistently outperforms existing SEMs by a notable margin, with 4.2%-8.7% relative gains in 5-way 5-shot classification.
  arXiv  Detail & Related papers  (2025-06-05T06:39:43Z)
• Interpreting the linear structure of vision-language model embedding spaces [12.846590038965774]
  We train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models.<n>SAEs approximate model embeddings as sparse linear combinations of learned directions, or "concepts"<n>Retraining SAEs with different seeds or different data diet leads to two findings: the rare, specific concepts captured by the SAEs are liable to change drastically, but we also show that commonly-activating concepts are remarkably stable across runs.
  arXiv  Detail & Related papers  (2025-04-16T01:40:06Z)
• LLM-based Hierarchical Concept Decomposition for Interpretable Fine-Grained Image Classification [5.8754760054410955]
  We introduce textttHi-CoDecomposition, a novel framework designed to enhance model interpretability through structured concept analysis. Our approach not only aligns with the performance of state-of-the-art models but also advances transparency by providing clear insights into the decision-making process.
  arXiv  Detail & Related papers  (2024-05-29T00:36:56Z)
• Understanding and Constructing Latent Modality Structures in Multi-modal Representation Learning [53.68371566336254]
  We argue that the key to better performance lies in meaningful latent modality structures instead of perfect modality alignment. Specifically, we design 1) a deep feature separation loss for intra-modality regularization; 2) a Brownian-bridge loss for inter-modality regularization; and 3) a geometric consistency loss for both intra- and inter-modality regularization.
  arXiv  Detail & Related papers  (2023-03-10T14:38:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.