Abstract

The quadratic complexity of Multimodal Large Language Models (MLLMs) with respect to sequence length poses significant computational and memory challenges, hindering their real-world deployment. While existing training-free token reduction methods aim to address these inefficiencies, how to precisely identify redundant visual tokens and recover the essential information from the discarded tokens remain unclear. In this paper, we propose a ''filter-correlate-compress'' framework that decomposes the token reduction into three stages: filtering redundant tokens, correlating discarded information to preserved tokens, and compressing tokens to minimize redundancy. Following the framework, we propose a solution FiCoCo to identify limitations in single redundancy assessment, propose adaptive strategies to retain critical information from discarded tokens, and mitigate semantic dilution during token fusion. Two specialized variants, FiCoCo-V (for vision encoders) and FiCoCo-L (for LLM decoders), further optimize efficiency across MLLM architectures. Extensive experiments demonstrate that FiCoCo achieves up to 5.7x/14.7x FLOPs reduction with 92.8%/93.6% performance retention on LLaVA-1.5-7B/LLaVA-NeXT-7B. Our methods consistently outperform state-of-the-art training-free approaches, showcasing effectiveness and generalizability across model architectures, sizes, and tasks without requiring retraining. Our project page is at https://ficoco-accelerate.github.io/.