MOSAIC: Adaptive Inter-layer Composition for Efficient Heterogeneous Vision-Language Models

• URL: http://arxiv.org/abs/2607.09029v1
• Date: Fri, 10 Jul 2026 01:31:02 GMT
• Title: MOSAIC: Adaptive Inter-layer Composition for Efficient Heterogeneous Vision-Language Models
• Abstract summary: Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data.<n>Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs.<n>We propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous operators.
• Score: 14.859177902733437
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs. However, these efforts rely on handcrafted static mixing patterns, which are sub-optimal and difficult to adapt to specific hardware. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms--including linear, sparse, and low-rank operators--into a unified search space. By formulating the selection as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate performance degradation from structural transitions, we introduce a two-stage parameter recovery process: global off-policy distillation to stabilize internal representations, followed by a dual-teacher on-policy distillation leveraging a 235B oracle for knowledge expansion and the original 4B teacher for distributional stability. We validate MOSAIC through MOSAIC-4B, derived from Qwen3-VL-4B-Instruct. Results demonstrate that MOSAIC-4B matches the baseline's performance across multiple benchmarks while requiring less than 2% of the original training cost. Furthermore, it substantially improves inference efficiency, achieving 1.76x prefilling and 2.54x decoding speedups.

Related papers

• AlignMamba-2: Enhancing Multimodal Fusion and Sentiment Analysis with Modality-Aware Mamba [56.52470564147458]
  We propose textbfAlignMamba-2, an effective and efficient framework for multimodal fusion and sentiment analysis.<n>Our approach introduces a dual alignment strategy that regularizes the model using both Optimal Transport distance and Maximum Mean Discrepancy.<n>More importantly, we design a Modality-Aware Mamba layer, which employs a Mixture-of-Experts architecture with modality-specific and modality-shared experts.
  arXiv  Detail & Related papers  (2026-03-19T03:47:21Z)
• DualGazeNet: A Biologically Inspired Dual-Gaze Query Network for Salient Object Detection [52.32976488996896]
  We introduce DualGazeNet, a pure Transformer framework for salient object detection.<n>Experiments on five RGB benchmarks show that DualGazeNet consistently surpasses 25 state-of-the-art CNN- and Transformer-based methods.
  arXiv  Detail & Related papers  (2025-11-24T08:08:22Z)
• MOBIUS: Big-to-Mobile Universal Instance Segmentation via Multi-modal Bottleneck Fusion and Calibrated Decoder Pruning [91.90342432541138]
  Scaling up model size and training data has advanced foundation models for instance-level perception.<n>High computational cost limits adoption on resource-constrained platforms.<n>We introduce a new benchmark for efficient segmentation on both high-performance computing platforms and mobile devices.
  arXiv  Detail & Related papers  (2025-10-16T18:00:00Z)
• OneCAT: Decoder-Only Auto-Regressive Model for Unified Understanding and Generation [91.45421429922506]
  OneCAT is a unified multimodal model that seamlessly integrates understanding, generation, and editing.<n>Our framework eliminates the need for external components such as Vision Transformers (ViT) or vision tokenizer during inference.
  arXiv  Detail & Related papers  (2025-09-03T17:29:50Z)
• LLM4CMO: Large Language Model-aided Algorithm Design for Constrained Multiobjective Optimization [54.35609820607923]
  Large language models (LLMs) offer new opportunities for assisting with algorithm design.<n>We propose LLM4CMO, a novel CMOEA based on a dual-population, two-stage framework.<n>LLMs can serve as efficient co-designers in the development of complex evolutionary optimization algorithms.
  arXiv  Detail & Related papers  (2025-08-16T02:00:57Z)
• Improving Progressive Generation with Decomposable Flow Matching [50.63174319509629]
  Decomposable Flow Matching (DFM) is a simple and effective framework for the progressive generation of visual media.<n>On Imagenet-1k 512px, DFM achieves 35.2% improvements in FDD scores over the base architecture and 26.4% over the best-performing baseline.
  arXiv  Detail & Related papers  (2025-06-24T17:58:02Z)
• HiLAB: A Hybrid Inverse-Design Framework [0.0]
  HiLAB is a new paradigm for inverse design of nanophotonic structures.<n>It addresses multi-functional device design by generating diverse freeform configurations at reduced simulation costs.
  arXiv  Detail & Related papers  (2025-05-23T05:34:56Z)
• HeteroTune: Efficient Federated Learning for Large Heterogeneous Models [35.53420882449293]
  We propose HeteroTune, a novel federated fine-tuning paradigm for large, heterogeneous models operating under limited communication and budgets.<n>The core of our method lies in a novel architecture, DeMA, which enables flexible and efficient aggregation of heterogeneous models.<n>We provide both theoretical analysis and empirical evidence showing that HeteroTune achieves state-of-the-art performance and efficiency across diverse tasks and model architectures.
  arXiv  Detail & Related papers  (2024-11-25T09:58:51Z)
• BiT-MamSleep: Bidirectional Temporal Mamba for EEG Sleep Staging [9.917709200378217]
  BiT-MamSleep is a novel architecture that integrates the Triple-Resolution CNN (TRCNN) for efficient multi-scale feature extraction. BiT-MamSleep incorporates an Adaptive Feature Recalibration (AFR) module and a temporal enhancement block to dynamically refine feature importance. Experiments on four public datasets demonstrate that BiT-MamSleep significantly outperforms state-of-the-art methods.
  arXiv  Detail & Related papers  (2024-11-03T14:49:11Z)
• Efficient Heterogeneous Large Language Model Decoding with Model-Attention Disaggregation [15.35494431928751]
  Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but also introduce significant challenges in real-world serving.<n>We introduce model-attention disaggregation to enhance the efficiency of LLM decoding.<n>We develop and deploy Lamina, an LLM inference system that incorporates model-attention disaggregation in a distributed heterogeneous cluster.
  arXiv  Detail & Related papers  (2024-05-03T02:15:15Z)
• Pruning Self-attentions into Convolutional Layers in Single Path [89.55361659622305]
  Vision Transformers (ViTs) have achieved impressive performance over various computer vision tasks. We propose Single-Path Vision Transformer pruning (SPViT) to efficiently and automatically compress the pre-trained ViTs. Our SPViT can trim 52.0% FLOPs for DeiT-B and get an impressive 0.6% top-1 accuracy gain simultaneously.
  arXiv  Detail & Related papers  (2021-11-23T11:35:54Z)

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.