How Visual Representations Map to Language Feature Space in Multimodal LLMs

• URL: http://arxiv.org/abs/2506.11976v2
• Date: Sun, 22 Jun 2025 00:39:23 GMT
• Title: How Visual Representations Map to Language Feature Space in Multimodal LLMs
• Authors: Constantin Venhoff, Ashkan Khakzar, Sonia Joseph, Philip Torr, Neel Nanda,
• Abstract summary: 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.
• Score: 9.880509106657009
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Effective multimodal reasoning depends on the alignment of visual and linguistic representations, yet the mechanisms by which vision-language models (VLMs) achieve this alignment remain poorly understood. Following the LiMBeR framework, we deliberately maintain a frozen large language model (LLM) and a frozen vision transformer (ViT), connected solely by training a linear adapter during visual instruction tuning. By keeping the language model frozen, we ensure it maintains its original language representations without adaptation to visual data. Consequently, the linear adapter must map visual features directly into the LLM's existing representational space rather than allowing the language model to develop specialized visual understanding through fine-tuning. Our experimental design uniquely enables the use of pre-trained sparse autoencoders (SAEs) of the LLM as analytical probes. These SAEs remain perfectly aligned with the unchanged language model and serve as a snapshot of the learned language feature-representations. Through systematic analysis of SAE reconstruction error, sparsity patterns, and feature SAE descriptions, we reveal the layer-wise progression through which visual representations gradually align with language feature representations, converging in middle-to-later layers. This suggests a fundamental misalignment between ViT outputs and early LLM layers, raising important questions about whether current adapter-based architectures optimally facilitate cross-modal representation learning.

Related papers

• Interpretable Zero-Shot Learning with Locally-Aligned Vision-Language Model [56.573203512455706]
  Large-scale vision-language models (VLMs) have achieved remarkable success in zero-shot learning (ZSL) by leveraging large-scale visual-text pair datasets.<n>One approach to address this issue is to develop interpretable models by integrating language.<n>We propose LaZSL, a locally-aligned vision-language model for interpretable ZSL.
  arXiv  Detail & Related papers  (2025-06-30T13:14:46Z)
• Decoupled Visual Interpretation and Linguistic Reasoning for Math Problem Solving [57.22004912994658]
  Current large vision-language models (LVLMs) typically employ a connector module to link visual features with text embeddings of large language models (LLMs)<n>This paper proposes a paradigm shift: instead of training end-to-end vision-language reasoning models, we advocate for developing a decoupled reasoning framework.
  arXiv  Detail & Related papers  (2025-05-23T08:18:00Z)
• AlignVLM: Bridging Vision and Language Latent Spaces for Multimodal Understanding [63.09928907734156]
  AlignVLM is a vision-text alignment method that maps visual features to a weighted average of text embeddings.<n>Our experiments show that AlignVLM achieves state-of-the-art performance compared to prior alignment methods.
  arXiv  Detail & Related papers  (2025-02-03T13:34:51Z)
• VladVA: Discriminative Fine-tuning of LVLMs [67.14293827774827]
  Contrastively-trained Vision-Language Models (VLMs) like CLIP have become the de facto approach for discriminative vision-language representation learning.<n>We propose to combine "the best of both worlds": a new training approach for discriminative fine-tuning of LVLMs.
  arXiv  Detail & Related papers  (2024-12-05T17:54:27Z)
• Towards Interpreting Visual Information Processing in Vision-Language Models [24.51408101801313]
  Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images.<n>We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM.
  arXiv  Detail & Related papers  (2024-10-09T17:55:02Z)
• SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs [40.74693126923826]
  Multimodal Large Language Models (MLLMs) have recently demonstrated remarkable perceptual and reasoning abilities. Training adapters with image-level supervision often results in significant misalignment. We introduce Supervised Embedding Alignment (SEA), a token-level alignment method that leverages vision-language pre-trained models.
  arXiv  Detail & Related papers  (2024-08-21T17:58:02Z)
• X-Former: Unifying Contrastive and Reconstruction Learning for MLLMs [49.30255148577368]
  X-Former is a lightweight transformer module designed to exploit the complementary strengths of CL and MIM. X-Former first bootstraps vision-language representation learning and multimodal-to-multimodal generative learning from two frozen vision encoders. It further bootstraps vision-to-language generative learning from a frozen LLM to ensure visual features from X-Former can be interpreted by the LLM.
  arXiv  Detail & Related papers  (2024-07-18T18:39:54Z)
• Bridging Vision and Language Spaces with Assignment Prediction [47.04855334955006]
  VLAP is a novel approach that bridges pretrained vision models and large language models (LLMs) We harness well-established word embeddings to bridge two modality embedding spaces. VLAP achieves substantial improvements over the previous linear transformation-based approaches.
  arXiv  Detail & Related papers  (2024-04-15T10:04:15Z)
• Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization [52.935150075484074]
  We introduce a well-designed visual tokenizer to translate the non-linguistic image into a sequence of discrete tokens like a foreign language. The resulting visual tokens encompass high-level semantics worthy of a word and also support dynamic sequence length varying from the image. This unification empowers LaVIT to serve as an impressive generalist interface to understand and generate multi-modal content simultaneously.
  arXiv  Detail & Related papers  (2023-09-09T03:01:38Z)

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.