Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders

• URL: http://arxiv.org/abs/2606.21197v1
• Date: Fri, 19 Jun 2026 08:08:43 GMT
• Title: Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders
• Abstract summary: We propose a framework based on Sparse Autoencoders (SAEs) to extract and analyze visual, textual, and multimodal concepts from Vision Language Models (VLMs)<n>Our framework improves visual concept quality by up to 45% compared to existing SAE-based methods.
• Score: 9.716523835964045
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision Language Models (VLMs) have demonstrated impressive performance in tasks requiring joint understanding of images and text, such as image captioning and Visual Question Answering (VQA), but our understanding of their internal processes remains limited. Recently, Sparse Autoencoders (SAEs) have emerged as a promising tool to support the interpretation of concepts encoded in VLMs. However, most SAE-based approaches focus only on textual or visual concepts separately, ignoring multimodal concepts. This limitation hinders a comprehensive understanding of VLMs, since concepts that integrate both modalities can be misclassified. Moreover, previous visual approaches often produce low-quality visual concept descriptions that are vague or incomplete, limiting their usefulness for understanding model reasoning. We propose a framework based on SAEs to extract and analyze visual, textual, and multimodal concepts from VLMs. For each neuron, we propose a candidate human-interpretable concept and compute the alignment between the concept and the dataset samples using cosine similarity scores. Experiments on a VQA dataset (LLaVA-NeXT) demonstrate that our framework improves visual concept quality by up to 45\% compared to existing SAE-based methods, while maintaining high textual concept quality and enabling systematic identification of multimodal concepts. This work contributes new insights into the conceptual space of VLMs, providing a structured approach to distinguish between visual, textual, and multimodal concepts. The code is available at https://github.com/PHDLanza/Multidata_SAE

Related papers

• 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)
• CoPA: Hierarchical Concept Prompting and Aggregating Network for Explainable Diagnosis [8.56688324078793]
  Concept Prompting and Aggregating (CoPA) is a novel framework designed to capture multilayer concepts under prompt guidance.<n>Visual representations from each layer are aggregated to align with textual concept representations.<n>CoPA outperforms state-of-the-art methods on three public datasets.
  arXiv  Detail & Related papers  (2025-10-04T10:29:15Z)
• MIMIC: Multimodal Inversion for Model Interpretation and Conceptualization [52.66401137323065]
  We propose a Multimodal Inversion for Model Interpretation and Conceptualization (MIMIC) framework to visualize the internal representations of Vision Language Models (VLMs)<n>MIMIC uses a joint VLM-based inversion and a feature alignment objective to account for VLM's autoregressive processing.<n>We quantitatively and qualitatively evaluate MIMIC by inverting visual concepts over a range of varying-length free-form VLM output texts.
  arXiv  Detail & Related papers  (2025-08-11T10:36:58Z)
• MC-LLaVA: Multi-Concept Personalized Vision-Language Model [51.645660375766575]
  This paper proposes the first multi-concept personalization paradigm, MC-LLaVA.<n>MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step.<n> Comprehensive qualitative and quantitative experiments demonstrate that MC-LLaVA can achieve impressive multi-concept personalized responses.
  arXiv  Detail & Related papers  (2025-03-24T16:32:17Z)
• CBVLM: Training-free Explainable Concept-based Large Vision Language Models for Medical Image Classification [7.641602003136712]
  Concept Bottleneck Models (CBMs) tackle the latter by constraining the model output on a set of predefined and human-interpretable concepts.<n>We propose a simple, yet effective, methodology, CBVLM, which tackles both of the aforementioned challenges.<n>We validate our approach with extensive experiments across four medical datasets and twelve LVLMs.
  arXiv  Detail & Related papers  (2025-01-21T16:38:04Z)
• V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer [19.177297480709512]
  Concept Bottleneck Models (CBMs) offer inherent interpretability by translating images into human-comprehensible concepts.<n>Recent approaches have leveraged the knowledge of large language models to construct concept bottlenecks.<n>In this study, we investigate to avoid these issues by constructing CBMs directly from multimodal models.
  arXiv  Detail & Related papers  (2025-01-09T05:12:38Z)
• MC-LLaVA: Multi-Concept Personalized Vision-Language Model [51.645660375766575]
  This paper proposes the first multi-concept personalization paradigm, MC-LLaVA.<n>MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step.<n> Comprehensive qualitative and quantitative experiments demonstrate that MC-LLaVA can achieve impressive multi-concept personalized responses.
  arXiv  Detail & Related papers  (2024-11-18T16:33:52Z)
• A Concept-Based Explainability Framework for Large Multimodal Models [52.37626977572413]
  We propose a dictionary learning based approach, applied to the representation of tokens.<n>We show that these concepts are well semantically grounded in both vision and text.<n>We show that the extracted multimodal concepts are useful to interpret representations of test samples.
  arXiv  Detail & Related papers  (2024-06-12T10:48:53Z)
• Pre-trained Vision-Language Models Learn Discoverable Visual Concepts [33.302556000017844]
  We aim to answer this question as visual concepts learned "for free" would enable wide applications.<n>We assume that the visual concepts, if captured by pre-trained VLMs, can be extracted by their vision-language interface with text-based concept prompts.<n>Our proposed concept discovery and learning framework is thus designed to identify a diverse list of generic visual concepts.
  arXiv  Detail & Related papers  (2024-04-19T06:41:32Z)
• MyVLM: Personalizing VLMs for User-Specific Queries [78.33252556805931]
  We take a first step toward the personalization of vision-language models, enabling them to learn and reason over user-provided concepts. To effectively recognize a variety of user-specific concepts, we augment the VLM with external concept heads that function as toggles for the model. Having recognized the concept, we learn a new concept embedding in the intermediate feature space of the VLM. This embedding is tasked with guiding the language model to naturally integrate the target concept in its generated response.
  arXiv  Detail & Related papers  (2024-03-21T17:51:01Z)

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.