Towards Efficient Multimodal Large Language Models: A Survey on Token Compression

Multimodal Large Language Models (MLLMs) have made significant strides in integrating vision-language perception, alignment, and reasoning. However, the increasing complexity of tasks such as high-resolution image processing and long video understanding has led to an exponential rise in visual context length within MLLMs. The resulting long-context token sequences impose substantial computational demands on large language models (LLMs), leading to quadratic complexity growth, heightened GPU resource consumption, and slower inference speeds. To address these challenges, token compression has emerged as a promising research direction that reduces the number of tokens processed within MLLMs while preserving essential cross-modal semantic information, thereby enhancing both training and inference efficiency. This survey provides a comprehensive review of token compression techniques for MLLMs, examining the current state of research and exploring future directions. We propose a taxonomy of token compression methods based on their application modules within the MLLM system, including the vision encoder, projector, LLM backbone, and hybrid approaches. We analyze the strengths and limitations of widely adopted algorithms, offering practitioners a structured framework for selecting appropriate token compression strategies. Finally, we discuss practical applications of token compression, identify key challenges in the field, and propose potential directions for future research and development. All related resources are available at https://github.com/yaolinli/MLLM-Token-Compression.