Principles and Methods for Building Efficient Large Vision Language Models

Large Vision–Language Models (LVLMs) have emerged as a powerful foundation for multimodal artificial intelligence, enabling unified reasoning over visual and textual information and achieving remarkable performance across a wide range of tasks, including image and video understanding, visual question answering, multimodal dialogue, document analysis, and embodied decision making. However, these capabilities have largely been driven by aggressive scaling in model size, training data, and computational resources, resulting in LVLMs that are costly to train, expensive to deploy, and difficult to scale in practical, resourceconstrained environments. As LVLMs transition from research prototypes to real-world systems, efficiency has become a critical bottleneck that fundamentally shapes their usability, accessibility, and sustainability. This survey provides a comprehensive and systematic review of research on efficient Large Vision–Language Models. We first establish a unified conceptual framework for understanding efficiency in the multimodal setting, encompassing computational complexity, memory footprint, inference latency, energy consumption, and data efficiency. Building on this foundation, we categorize and analyze existing approaches to efficiency from multiple perspectives, including architectural design, training and optimization strategies, parameter-efficient adaptation, model compression, and inference acceleration. We highlight how modality-aware architectures, lightweight multimodal interfaces, conditional computation, and token-level optimization can substantially reduce redundancy while preserving expressive power. In parallel, we examine training paradigms such as decoupled pretraining, knowledge distillation, and curriculum learning that improve convergence speed and performance under limited resources. We further review post-training techniques such as pruning, quantization, low-rank factorization, and hardware-aware optimization, emphasizing their role in enabling fast and memory-efficient deployment of LVLMs across diverse platforms, from cloud-scale systems to edge and on-device applications. In addition to summarizing representative methods, we discuss commonly used benchmarks, evaluation protocols, and efficiency metrics, and analyze the practical trade-offs that arise when balancing accuracy and resource constraints. Beyond existing techniques, this survey identifies key open challenges and future research directions, including standardized efficiency evaluation, long-context multimodal reasoning, robustness under aggressive compression, model–hardware co-design, and the environmental and societal implications of large-scale multimodal models. By synthesizing a rapidly growing and fragmented body of literature, this survey aims to provide researchers and practitioners with a clear roadmap for developing scalable, deployable, and sustainable LVLMs, and to position efficiency as a first-class objective in the next generation of multimodal intelligence systems.