Efficient Architectures and Learning Strategies for Large Vision Language Models

Large Vision-Language Models (LVLMs) have emerged as a powerful paradigm for multimodal artificial intelligence, enabling unified reasoning over visual and textual inputs and achieving remarkable performance across a wide range of tasks, including image captioning, visual question answering, multimodal dialogue, document understanding, and embodied decision-making. By coupling high-capacity visual encoders with large-scale language models, LVLMs exhibit strong generalization and zero-shot capabilities, positioning them as foundational components of next-generation AI systems. However, these advances have been accompanied by rapidly increasing computational, memory, and energy costs, which pose significant barriers to scalable training, efficient inference, and real-world deployment. As LVLMs continue to grow in size and scope, efficiency has become a central concern that fundamentally shapes model design, learning strategies, and evaluation practices. This survey provides a comprehensive and systematic review of research on efficient Large Vision-Language Models. We present a unified perspective on efficiency that encompasses architectural design, parameterefficient training and adaptation, model compression and inference acceleration, data-efficient learning, and training-efficient optimization strategies. We begin by formalizing the LVLM problem and introducing commonly used efficiency metrics, including parameter count, floating-point operations, memory footprint, latency, and energy consumption, highlighting the inherent trade-offs between multimodal performance and resource usage. Building on this foundation, we examine architectural approaches that reduce computational redundancy in visual encoding and multimodal fusion, such as lightweight vision backbones, bottleneckbased cross-modal interaction, and modular decoupling of vision and language components. We further survey parameter-efficient adaptation techniques, including adapters, low-rank updates, and prompt-based methods, which enable LVLMs to be specialized for downstream tasks while updating only a small fraction of parameters. Model compression methods-such as pruning, quantization, and knowledge distillation-are reviewed in the context of multimodal systems, with particular attention to their impact on inference latency and robustness. In addition, we discuss data-efficient and training-efficient strategies that reduce dependence on large-scale labeled datasets and long training cycles, including self-supervised pretraining, curriculum learning, synthetic data generation, and staged training pipelines. Beyond summarizing existing methods, this survey critically analyzes evaluation protocols and benchmarks for efficient LVLMs, emphasizing the need for standardized, multidimensional, and deployment-aware assessment practices. We identify key open challenges and emerging directions, such as adaptive computation, hardware-aware co-design, efficient video and embodied LVLMs, and sustainability-oriented evaluation. By synthesizing a rapidly growing and fragmented body of literature, this survey aims to serve as a comprehensive reference for researchers and practitioners, and to provide conceptual guidance for the development of scalable, practical, and sustainable Large Vision–Language Models.