Vision Function Layer in Multimodal LLMs

• URL: http://arxiv.org/abs/2509.24791v1
• Date: Mon, 29 Sep 2025 13:45:35 GMT
• Title: Vision Function Layer in Multimodal LLMs
• Authors: Cheng Shi, Yizhou Yu, Sibei Yang,
• Abstract summary: This study identifies visual-related functional decoding is distributed across different decoder layers in Multimodal Large Language Models (MLLMs)<n>Visual Token Swapping modifies targeted KV cache entries to precisely elucidate layer-specific functions during decoding.<n>This study delivers deeper comprehension of MLLM visual processing, fostering the creation of more efficient, interpretable, and robust models.
• Score: 70.42157905484765
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: This study identifies that visual-related functional decoding is distributed across different decoder layers in Multimodal Large Language Models (MLLMs). Typically, each function, such as counting, grounding, or OCR recognition, narrows down to two or three layers, which we define as Vision Function Layers (VFL). Additionally, the depth and its order of different VFLs exhibits a consistent pattern across different MLLMs, which is well-aligned with human behaviors (e.g., recognition occurs first, followed by counting, and then grounding). These findings are derived from Visual Token Swapping, our novel analytical framework that modifies targeted KV cache entries to precisely elucidate layer-specific functions during decoding. Furthermore, these insights offer substantial utility in tailoring MLLMs for real-world downstream applications. For instance, when LoRA training is selectively applied to VFLs whose functions align with the training data, VFL-LoRA not only outperform full-LoRA but also prevent out-of-domain function forgetting. Moreover, by analyzing the performance differential on training data when particular VFLs are ablated, VFL-select automatically classifies data by function, enabling highly efficient data selection to directly bolster corresponding capabilities. Consequently, VFL-select surpasses human experts in data selection, and achieves 98% of full-data performance with only 20% of the original dataset. This study delivers deeper comprehension of MLLM visual processing, fostering the creation of more efficient, interpretable, and robust models.

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)
• A Comprehensive Study on Visual Token Redundancy for Discrete Diffusion-based Multimodal Large Language Models [85.30893355216486]
  We study how visual token redundancy evolves with different dMLLM architectures and tasks.<n>Our study reveals that visual redundancy emerges only in from-scratch dMLLMs while handling long-answer tasks.<n>Layer-skipping is promising for accelerating AR-to-diffusion dMLLMs, whereas progressive or late-step pruning is more effective for from-scratch dMLLMs.
  arXiv  Detail & Related papers  (2025-11-19T04:13:36Z)
• RL makes MLLMs see better than SFT [96.508432109136]
  We conduct a critical yet under-explored analysis of the vision encoder of Multimodal Language Model (MLLM)<n>Our results demonstrate that MLLM's post-training strategy (i.e., SFT or RL) not only leads to distinct outcomes on MLLM downstream tasks, but also fundamentally reshapes MLLM's underlying visual representations.<n>We then reframe our findings into a simple recipe for building strong vision encoders for MLLMs, Preference-Instructed Vision OpTimization (PIVOT)
  arXiv  Detail & Related papers  (2025-10-18T03:37:17Z)
• Dynamic Embedding of Hierarchical Visual Features for Efficient Vision-Language Fine-Tuning [5.85033069870214]
  We propose an efficient vision-language fine-tuning method based on dynamic embedding and fusion of hierarchical visual features.<n>By fine-tuning only a small number of parameters, DEHVF achieves precise alignment and complement of cross-modal information.
  arXiv  Detail & Related papers  (2025-08-25T03:57:46Z)
• LLaVA-UHD v2: an MLLM Integrating High-Resolution Semantic Pyramid via Hierarchical Window Transformer [110.39467860530819]
  Vision transformers (ViTs) are widely employed in multimodal large language models (MLLMs) for visual encoding.<n>We present LLaVA-UHD v2, an MLLM with advanced perception abilities by introducing a well-designed vision-language projector.<n>Hiwin transformer enhances MLLM's ability to capture diverse multi-modal visual granularities, by incorporating our constructed high-resolution semantic pyramid.
  arXiv  Detail & Related papers  (2024-12-18T14:07:46Z)
• Elevating Visual Perception in Multimodal LLMs with Auxiliary Embedding Distillation [109.5893580175657]
  In recent times, the standard practice for developing MLLMs is to feed features from vision encoder(s) into the LLM and train with natural language supervision.<n>This approach often causes models to lean towards language comprehension and undermine the rich visual perception signals present in the data.<n>We propose VisPer-LM, the first approach that infuses visual perception knowledge from expert vision encoders into the LLM's hidden representations.
  arXiv  Detail & Related papers  (2024-12-12T18:55:18Z)
• 60 Data Points are Sufficient to Fine-Tune LLMs for Question-Answering [50.12622877002846]
  Large language models (LLMs) encode extensive world knowledge through pre-training on massive datasets, which can be fine-tuned for the question-answering (QA) task.<n>We categorize supervised fine-tuning (SFT) data based on the extent of knowledge memorized by the pretrained LLMs.<n>Our experiments show that as few as 60 data points during the SFT stage can activate the knowledge encoded during pre-training, enabling LLMs to perform the QA task.
  arXiv  Detail & Related papers  (2024-09-24T07:38:38Z)
• Bridge the Modality and Capability Gaps in Vision-Language Model Selection [62.26769826687365]
  Vision Language Models (VLMs) excel in zero-shot image classification by pairing images with textual category names.<n>To better reuse the VLM resource, a promising strategy is selecting appropriate Pre-Trained VLMs from the VLM Zoo.<n>We analyze two inherent challenges in assessing the ability of a VLM in this Language-Only VLM selection.<n>We propose VLM Selection With gAp Bridging to mitigate the negative impact of two gaps.
  arXiv  Detail & Related papers  (2024-03-20T17:54:58Z)
• VertiBench: Advancing Feature Distribution Diversity in Vertical Federated Learning Benchmarks [31.08004805380727]
  This paper introduces two key factors affecting VFL performance - feature importance and feature correlation. We also introduce a real VFL dataset to address the deficit in image-image VFL scenarios.
  arXiv  Detail & Related papers  (2023-07-05T05:55:08Z)
• BlindFL: Vertical Federated Machine Learning without Peeking into Your Data [20.048695060411774]
  Vertical federated learning (VFL) describes a case where ML models are built upon the private data of different participated parties. We introduce BlindFL, a novel framework for VFL training and inference. We show that BlindFL supports diverse datasets and models efficiently whilst achieving robust privacy guarantees.
  arXiv  Detail & Related papers  (2022-06-16T07:26:50Z)

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.