Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper analyzes the inefficiencies in multimodal large language models (MLLMs) that stem from applying a uniform Transformer architecture to both image and text tokens, revealing that vision tokens saturate in the middle layers while text tokens benefit from deeper processing. The authors introduce Dual-Path Vision Token Routing (DPVR), specifically its Late-Layer Fusion (DPVR-LF) variant, which routes vision tokens into a separate trainable branch and only fuses them with text tokens at the final layer. This approach reduces visual computation while maintaining competitive performance, demonstrating that a single late fusion layer can effectively support multimodal tasks without the need for extensive visual processing through all layers.
A single late fusion layer is enough to maintain multimodal performance, challenging the need for vision tokens to traverse all layers of a Transformer.
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.