A Structured Sparse AutoEncoder that enforces concept consistency from both semantic and spatial perspectives in the visual modality is proposed and introduced, and enhances neuronal monosemanticity by this visual structural prior.
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.
It is found that SAE activation sets do not recover human category boundaries or within-category typicality more faithfully than dense embeddings or residual-stream states, but instead track model-internal similarity structure.
Nikolai Bolik, Lennart Stöpler, Artur Andrzejak· 0 citations
Mechanistic interpretability has recently expanded to Vision Transformers (ViTs), with Sparse Autoencoders (SAEs) increasingly used as post-hoc tools to decompose internal representations into sparse and more interpretable features. However, because post-hoc SAEs are trained on frozen representations after the ViT has already been optimized, their latent features are not directly aligned with the downstream classification objective. We introduce SpIn-ViT, a framework that jointly trains a pretrained ViT and a modified SAE end-to-end, directly aligning sparse patch-level representations with image classification. SpIn-ViT learns semantically coherent neuron activations that localize meaningful image regions while maintaining competitive predictive performance. We evaluate SpIn-ViT across nine image-classification benchmarks using classification accuracy, quantitative interpretability metrics, AI-based and Human evaluations. Compared with the previous state-of-the-art post-hoc SAE method, SpIn-ViT achieves 8.84% higher average classification accuracy, an AI-based interpretability score nearly four times as high, and a human-evaluation score more than twice as high. We further extract interpretable rule-sets using the SAE neurons to create neurosymbolic models which achieve 5.97% higher average classification accuracy while requiring a 58.8\% smaller rule-set than the neurosymbolic models created from the SOTA post-hoc SAE method.
Unified multimodal models seek a shared visual token space that supports both multimodal understanding and image generation. Discrete methods unify the interface via a shared codebook, whereas continuous pipelines often rely on two disparate representations -- semantic features (e.g., ViT) for understanding and low-level latents (e.g., VAE) for synthesis -- resulting in mismatched latent spaces. We propose Twins, a unified continuous token space formed by channel-wise concatenating ViT and VAE features on the same token grid, so the sequence length is unchanged and attention cost does not increase. However, jointly modeling Twins in a Diffusion Transformer exposes a severe optimization imbalance: the model fits the ViT component well but struggles to match the VAE latent distribution. We trace this imbalance to three sources of heterogeneity: frequency bias, intrinsic dimensionality, and condition-aligned vs condition-independent uncertainty. To address it, we adapt a focal regression objective for flow matching that upweights large-error VAE dimensions, better balancing optimization across the ViT and VAE components. On ImageNet, this yields up to 10.57 gFID gain over naive MSE loss without classifier-free guidance. Twins also performs competitively on multimodal understanding benchmarks and improves reconstruction fidelity, narrowing the gap between understanding- and generation-oriented representations.
Kaixiong Gong, Xin Cai, Bin Lin et al.· 0 citations
Learning disentangled representations is a key requirement for developing versatile, general-purpose, and sustainable models in multi-modal wearable computing. However, existing approaches do not operate as full-stack wearable processors, i.e., they do not simultaneously address task-specific classification performance, disentangled and interpretable representation learning, fusion, and generative modeling of highly heterogeneous multi-modal time series. To address this gap, we introduce Omni-modal Variational Decomposition Autoencoders (OmniDecVAEs), a framework that efficiently learns multi-purpose representations in a unified and scalable manner from arbitrarily many modalities. OmniDecVAEs extend DecVAEs by learning modality-conditioned time-frequency latent subspaces through a multi-view self-supervised decomposition loss and a shared asymmetric autoencoder (AE) architecture. Results on a challenging omni-modal human activity recognition (HAR) setting with up to thirty modalities, demonstrate the ability of OmniDecVAEs to learn full-stack wearable representations. When compared to transformer-based and VAE-based methods, OmniDecVAEs full-stack disentangled representation properties lead to accuracy improvements of 1.01% and 6.75% in activity and identity recognition, respectively. Furthermore, OmniDecVAEs synthesize realistic omni-modal time-frequency data that manifest with enhanced reconstructions (mean absolute error improves by 76.84%) and distributional similarity between real and synthetic data (maximum mean discrepancy improves by 13.85%). Our results highlight OmniDecVAEs potential as a lightweight model suitable for intelligent edge wearables and clinical healthcare, unifying processing requirements and abilities in a single model, through its enhanced representational capacity, modality-invariant spatial complexity (4.1M parameters), and real-time latency.
Ioannis Ziogas, Ensieh Khazaei, Bilal Taha et al.· 0 citations
Leveraging textual information for image clustering has emerged as a promising direction, largely owing to the powerful representations learned by Vision-Language Models (VLMs). Existing approaches typically retrieve a textual counterpart for each image and then refine multimodal representations by directly enforcing cross-modal agreement, e.g., maximizing image-text similarity inherited from pretrained VLMs. However, such a strategy aligns heterogeneous representations across modalities without explicitly modeling the intrinsic structure within each modality and thus might yield unreliable alignment or distort modality-specific structures that are crucial for clustering. In this paper, we propose a simple but principled approach, termed deep modality-shared self-expressive model (DeepMORSE), which discovers cross-modal structures via a modality-shared self-expressive model and simultaneously learns structured representations that conform to a union of modality-specific subspaces. Moreover, we theoretically justify that the modality-shared self-expressive coefficients suppress inter-class noise towards a subspace-preserving solution, and show that mini-batch optimization procedure introduces an implicit regularization onto the self-expressive model. We evaluate our DeepMORSE on six widely used image clustering benchmarks and observe performance improvements exceeding 3% on the UCF-101, DTD-47, and ImageNet-Dogs datasets. In addition, we demonstrate the strong transferability of the learned representations by achieving state-of-the-art performance on downstream tasks such as image retrieval and zero-shot classification---without requiring any task-specific losses or post-processing. The code is available at: https://github.com/mengxianghan123/DeepMORSE.
Xianghan Meng, Wei He, Zhiyuan Huang et al.· 0 citations
Within Explainable Artificial Intelligence, mechanistic interpretability uses Sparse Autoencoders (SAEs) to extract more interpretable features from neural representations. However, assessing their monosemanticity, and thus explanation quality, remains challenging. Existing metrics require external concept labels or depend on pretrained embedding models, making them sensitive to encoder's geometry. We introduce the Tversky Monosemanticity Score (TMS), a label-free metric that operationalizes monosemanticity as activation-set coherence of binarized SAE latents, and does not require external embedding encoders. We evaluate TMS on SAEs trained on features from pretrained vision and vision-language models (DINOv3, CLIP, BLIP2), two common SAE regimes (TopK, BatchTopK), multiple sparsity levels, and expansion factors. Our results show that TMS is less affected by encoder anisotropy than its embedding-based alternative, while remaining aligned with established monosemanticity indicators. TMS also reveals distinct SAE training dynamics across base models. Moreover, under encoder anisotropy, TMS provides a stronger indication of probe-based concept deletion effectiveness, while being competitive otherwise.
Katarzyna Filus, Sebastian Pokuciński· 0 citations
