Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs

• URL: http://arxiv.org/abs/2606.16193v1
• Date: Mon, 15 Jun 2026 04:10:40 GMT
• Title: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs
• Abstract summary: We introduce cascaded autoencoders (CSAEs) for learning hierarchical visual concepts in MLLMs.<n> experiments across Qwen3-VL, Gemma-3, and LLaVA on multiple visual datasets show that CSAEs improve interpretability in terms of hierarchical concept coherence.<n>Results on concept steering further demonstrate that the learned concept groups support effective group-level interventions in MLLM outputs.
• Score: 12.597421665142646
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representations remain difficult to interpret. Sparse Autoencoders (SAEs) provide a scalable way to decompose dense model activations into sparse, interpretable features. However, existing SAE architectures primarily recover flat feature dictionaries and are less suited for explicit multi-level concept organization. In this paper, we introduce cascaded sparse autoencoders (CSAEs) for learning hierarchical visual concepts in MLLMs. Rather than nesting or stacking SAE sparse activation codes, CSAEs train a second-level SAE directly on the decoder weights of the first-level SAE, treating learned low-level feature directions as inputs for higher-level abstraction. This design enables CSAEs to learn "concepts of concepts" while avoiding drawbacks from the shared-prefix coupling of nesting, Matryoshka-style hierarchies and the bottlenecks of naively stacked SAEs. Experiments across Qwen3-VL, Gemma-3, and LLaVA on multiple visual datasets show that CSAEs improve interpretability in terms of hierarchical concept coherence over state-of-the-art SAE baselines. Results on concept steering further demonstrate that the learned concept groups support effective group-level interventions in MLLM outputs.

Related papers

• From One-to-One to Many-to-Many: Dynamic Cross-Layer Injection for Deep Vision-Language Fusion [91.35078719566472]
  Vision-Language Models (VLMs) create a severe visual feature bottleneck by using a crude, asymmetric connection.<n>We introduce Cross-Layer Injection (CLI), a novel and lightweight framework that forges a dynamic many-to-many bridge between the two modalities.
  arXiv  Detail & Related papers  (2026-01-15T18:59:10Z)
• Seeing Beyond Words: Self-Supervised Visual Learning for Multimodal Large Language Models [53.06230963851451]
  JARVIS is a JEPA-inspired framework for self-supervised visual enhancement in MLLMs.<n>We introduce JARVIS, a JEPA-inspired framework for self-supervised visual enhancement in MLLMs.
  arXiv  Detail & Related papers  (2025-12-17T19:01:34Z)
• LISA: A Layer-wise Integration and Suppression Approach for Hallucination Mitigation in Multimodal Large Language Models [8.122679857175315]
  Multimodal Large Language Models (MLLMs) excel in vision-language tasks but remain prone to object hallucinations.<n>We propose textbfLISA, which enhances generation consistency through hierarchical modulation and multi-layer fusion.<n>Experiments show that LISA reduces hallucinations by up to 53.6% in $mathrmCHAIR_I$ and improves POPE F1 by 4.5%.
  arXiv  Detail & Related papers  (2025-07-25T09:48:23Z)
• How Visual Representations Map to Language Feature Space in Multimodal LLMs [9.880509106657009]
  We study the mechanism by which vision-language models (VLMs) achieve alignment of visual and linguistic representations.<n>By keeping the language model frozen, we ensure it maintains its original language representations without adaptation to visual data.<n>We reveal the layer-wise progression through which visual representations gradually align with language feature representations, converging in middle-to-later layers.
  arXiv  Detail & Related papers  (2025-06-13T17:34:05Z)
• Abstractive Visual Understanding of Multi-modal Structured Knowledge: A New Perspective for MLLM Evaluation [48.462734327375536]
  Multi-modal large language models (MLLMs) incorporate heterogeneous modalities into LLMs, enabling a comprehensive understanding of diverse scenarios and objects.<n>Despite the proliferation of evaluation benchmarks and leaderboards for MLLMs, they predominantly overlook the critical capacity of MLLMs to comprehend world knowledge with structured abstractions that appear in visual form.<n>We propose M3STR, an innovative benchmark grounded in the Multi-Modal Map for STRuctured understanding.<n>Our findings reveal persistent deficiencies in processing abstractive visual information with structured knowledge, thereby charting a pivotal trajectory for advancing MLLMs' holistic reasoning capacities.
  arXiv  Detail & Related papers  (2025-06-02T04:00:35Z)
• Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models [50.587868616659826]
  We introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in vision representations.<n>Our experimental results reveal that SAEs trained on Vision-Language Models significantly enhance the monosemanticity of individual neurons.
  arXiv  Detail & Related papers  (2025-04-03T17:58:35Z)
• Instruction-Guided Fusion of Multi-Layer Visual Features in Large Vision-Language Models [50.98559225639266]
  We investigate the contributions of visual features from different encoder layers using 18 benchmarks spanning 6 task categories.<n>Our findings reveal that multilayer features provide complementary strengths with varying task dependencies, and uniform fusion leads to suboptimal performance.<n>We propose the instruction-guided vision aggregator, a module that dynamically integrates multi-layer visual features based on textual instructions.
  arXiv  Detail & Related papers  (2024-12-26T05:41:31Z)
• SynerGen-VL: Towards Synergistic Image Understanding and Generation with Vision Experts and Token Folding [66.74446220401296]
  We propose SynerGen-VL, a simple yet powerful encoder-free MLLM capable of both image understanding and generation.<n>We introduce the token folding mechanism and the vision-expert-based progressive alignment pretraining strategy, which effectively support high-resolution image understanding.<n>Our code and models shall be released.
  arXiv  Detail & Related papers  (2024-12-12T18:59:26Z)
• SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs [31.88022265176855]
  Supervised Embedding Alignment (SEA) is a token-level supervision alignment method that enables more precise visual-text alignment during pretraining.<n>Our comprehensive analyses reveal critical insights into the adapter's role in multimodal integration, and extensive experiments demonstrate that SEA consistently improves performance across various model sizes.
  arXiv  Detail & Related papers  (2024-08-21T17:58:02Z)
• LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge [58.82222646803248]
  Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. Most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. We propose a dual-Level vIsual knedgeOwl eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels.
  arXiv  Detail & Related papers  (2023-11-20T15:56:44Z)
• From CLIP to DINO: Visual Encoders Shout in Multi-modal Large Language Models [36.41816380074965]
  We investigate the effectiveness of different vision encoders within Large Language Models (MLLMs) Our findings reveal that the shallow layer features of CLIP offer particular advantages for fine-grained tasks such as grounding and region understanding. We propose a simple yet effective feature merging strategy, named COMM, that integrates CLIP and DINO with Multi-level features Merging.
  arXiv  Detail & Related papers  (2023-10-13T02:41:55Z)

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.