Towards Computationally Efficient Large Vision Language Models for Scalable Multimodal Intelligence

Large Vision-Language Models (LVLMs) have emerged as a powerful paradigm for multimodal artificial intelligence, enabling unified systems that jointly process visual and linguistic information to perform a wide range of tasks such as image captioning, visual question answering, multimodal reasoning, document understanding, and instruction following. By combining large-scale vision encoders with pretrained large language models, LVLMs have achieved unprecedented levels of performance and generalization. However, these advances have come at the cost of substantial computational, memory, and energy requirements, posing significant challenges for training, inference, and real-world deployment. As LVLMs continue to scale in size and scope, efficiency has become a critical concern that directly impacts their accessibility, sustainability, and practical applicability. This survey provides a comprehensive and systematic review of research on efficient Large Vision-Language Models. We analyze efficiency as a multi-dimensional concept encompassing model size, computational complexity, memory footprint, latency, energy consumption, and system-level scalability. We first formalize the vision-language learning problem and the canonical architectures underlying modern LVLMs, highlighting the primary sources of computational inefficiency. We then introduce a unified taxonomy that categorizes existing efficiency techniques across the entire LVLM lifecycle, including data-centric methods, architectural design choices, training-time optimization strategies, inference-time acceleration, and system-level and deployment-oriented techniques We further examine commonly used benchmarks, evaluation protocols, and efficiency metrics, discussing their strengths, limitations, and implications for fair comparison. Through a comparative analysis of representative efficient LVLMs, we illustrate how different design philosophies lead to distinct efficiency–performance trade-offs, and why no single approach is universally optimal. Finally, we identify key open challenges and future research directions, such as efficient multimodal reasoning, scalable video-language modeling, continual learning under resource constraints, robustness-aware efficiency optimization, and hardware–algorithm co-design. By synthesizing insights from a rapidly growing and fragmented body of literature, this survey aims to provide a clear and structured reference for researchers and practitioners interested in building efficient, scalable, and practical vision-language systems. We hope that this work will facilitate principled design choices, promote standardized evaluation practices, and ultimately accelerate the development and deployment of LVLMs that are not only powerful, but also efficient, sustainable, and broadly accessible