Late-Layer Fusion is Enough: Dual-Path Vision Token Routing for Multimodal Large Language Models under Visual Saturation

• URL: http://arxiv.org/abs/2606.09131v1
• Date: Mon, 08 Jun 2026 07:28:14 GMT
• Title: Late-Layer Fusion is Enough: Dual-Path Vision Token Routing for Multimodal Large Language Models under Visual Saturation
• Abstract summary: Multimodal large language models (MLLMs) commonly inherit the deep, symmetric Transformer backbone designed for unimodal text modeling.<n>This design overlooks a key modality asymmetry: image and text tokens differ substantially in information density, redundancy, and required reasoning depth.<n>We propose Dual-Path Vision Token Routing (DPVR), a modality-asymmetric routing framework for efficient MLLMs.
• Score: 6.369257323378483
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multimodal large language models (MLLMs) commonly inherit the deep, symmetric Transformer backbone designed for unimodal text modeling, and apply the same computation uniformly to image and language tokens. This design overlooks a key modality asymmetry: image and text tokens differ substantially in information density, redundancy, and required reasoning depth. Through a layer-wise analysis of LLaVA-1.5, we observe that vision tokens tend to saturate in the middle layers. Specifically, text-to-image attention decreases from 0.68 at layer 0 to 0.07 by layer 4, and stabilizes near 0.04 after layer 18, whereas text tokens continue to benefit from deep semantic processing. These findings suggest a mismatch between architectural symmetry and depth-asynchronous modality evolution, resulting in redundant visual computation and possible drift in perceptual representations during deep task-specific adaptation. Motivated by this, we propose Dual-Path Vision Token Routing (DPVR), a modality-asymmetric routing framework for efficient MLLMs. Its core instantiation, DPVR-LF (Late-Layer Fusion), routes vision tokens at the saturation point into a one-layer trainable side branch, runs a thirteen-layer text-only forward that skips image positions in the deep stack, and re-fuses the visual and textual streams only at the final layer. With approximately 3% trainable parameters, DPVR-LF preserves competitive multimodal performance on standard benchmarks while reducing visual computation in the deep Transformer stack. The results challenge the conventional assumption that vision tokens must traverse all deep language-model layers, and indicate that a single late fusion layer can be sufficient for maintaining strong perceptual competence in LLaVA-style MLLMs.

Related papers

• Beyond Text Conditioning: A Systematic Study of MLLM-DiT Fusion for Video Generation [50.92925099642139]
  Diffusion Transformers (DiTs) have become the dominant paradigm for high-fidelity video generation, yet their ability to perform high-level semantic planning remains limited.<n>We propose BiVidGen, a hybrid framework where an MLLM first generates semantic visual tokens and a DiT renders videos conditioned on both text and these tokens via multi-layer cross-attention.
  arXiv  Detail & Related papers  (2026-08-14T07:40:41Z)
• Decoupled Visual Processing: Efficient Multimodal Adaptation via Modality-Specific Transformer Substitution [26.422874365108985]
  We propose Decoupled Visual Processing (DVP) as an efficient training framework for multimodal large language models.<n>During training, only the single transformer block is updated, dramatically reducing the number of trainable parameters.<n>Experiments on the LLaVA-1.5 framework demonstrate that DVP competitive performance on MME, POPE, and ChartQA benchmarks while training only a fraction of the total parameters.
  arXiv  Detail & Related papers  (2026-07-29T08:16:28Z)
• Unlocking Dense Metric Depth Estimation in VLMs [36.431792667223796]
  Vision-Language Models (VLMs) excel at 2D tasks such as grounding and captioning, yet remain limited in 3D understanding.<n>We propose DepthVLM, a simple yet effective framework that transforms a single VLM into a native dense geometry predictor.<n>By attaching a lightweight depth head to the LLM backbone, DepthVLM generates full-resolution depth maps alongside language outputs in a single forward pass.
  arXiv  Detail & Related papers  (2026-05-15T11:54:17Z)
• Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation [66.53544128707817]
  Cheers is a unified multimodal model that decouples patch-level details from semantic representations.<n> Cheers matches or surpasses advanced UMMs in both visual understanding and generation.
  arXiv  Detail & Related papers  (2026-03-13T08:55:27Z)
• ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention [22.397648349603696]
  ViCA is a minimal MLLM architecture in which visual tokens bypass all self-attention and feed-forward layers, interacting with text solely through sparse cross-attention at selected layers.<n>ViCA preserves 98% of baseline accuracy while reducing visual-side to 4%, consistently achieving superior performance-efficiency trade-offs.
  arXiv  Detail & Related papers  (2026-02-07T14:46:05Z)
• ResTok: Learning Hierarchical Residuals in 1D Visual Tokenizers for Autoregressive Image Generation [64.84095852784714]
  Residual Tokenizer (ResTok) is a 1D visual tokenizer that builds hierarchical residuals for both image tokens and latent tokens.<n>We show that restoring hierarchical residual priors in visual tokenization significantly improves AR image generation, achieving a gFID of 2.34 on ImageNet-256 with only 9 sampling steps.
  arXiv  Detail & Related papers  (2026-01-07T14:09:18Z)
• Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials [34.77694214755808]
  We introduce Visual-Contrast Attention (VCA), a drop-in replacement for Multi-Head Self-Attention (MHSA)<n>VCA injects an explicit notion of discrimination while reducing the theoretical complexity from O(N N C) to O(N n C) with n N.<n>The module adds fewer than 0.3M parameters to a DeiT-Tiny backbone, requires no extra FLOPs, and is wholly architecture-agnostic.
  arXiv  Detail & Related papers  (2025-11-02T07:04:12Z)
• $\mathcal{V}isi\mathcal{P}runer$: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs [26.779915891040236]
  We propose emphVisiPruner, a training-free pruning framework that reduces up to 99% of vision-related attention computations and 53.9% of FLOPs on LLaVA-v1.5 7B.<n>Our insights further provide actionable guidelines for training efficient MLLMs by aligning model architecture with its intrinsic layer-wise processing dynamics.
  arXiv  Detail & Related papers  (2025-10-20T06:40:17Z)
• Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs [88.68484904214142]
  We introduce Patch-as-Decodable Token (PaDT) to generate both textual and diverse visual outputs.<n>Central to PaDT are Visual Reference Tokens (VRTs), derived from visual patch embeddings of query images.<n>We show PaDT consistently achieves state-of-the-art performance, even compared with significantly larger MLLM models.
  arXiv  Detail & Related papers  (2025-10-02T12:23:57Z)
• Token-Shuffle: Towards High-Resolution Image Generation with Autoregressive Models [92.18057318458528]
  Token-Shuffle is a novel method that reduces the number of image tokens in Transformer.<n>Our strategy requires no additional pretrained text-encoder and enables MLLMs to support extremely high-resolution image synthesis.<n>In GenAI-benchmark, our 2.7B model achieves 0.77 overall score on hard prompts, outperforming AR models LlamaGen by 0.18 and diffusion models LDM by 0.15.
  arXiv  Detail & Related papers  (2025-04-24T17:59:56Z)
• Planting a SEED of Vision in Large Language Model [73.17530130368053]
  We present SEED, an elaborate image tokenizer that empowers Large Language Models (LLMs) with the ability to SEE and Draw at the same time. This version of SEED was trained in 5.7 days using only 64 V100 GPUs and 5M publicly available image-text pairs.
  arXiv  Detail & Related papers  (2023-07-16T13:41:39Z)

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.