Fre-Res: A New Video Token Compression Framework for MLLMs

A new research paper presents Fre-Res, a cost-effective dual-track video-token compression system aimed at resolving the conflict between spatial accuracy and temporal coverage in Video Multimodal Large Language Models (MLLMs). This framework distinguishes between spatial and temporal data by maintaining sparse, high-fidelity spatial anchors while utilizing compact residual-frequency tokens to depict dense temporal changes. It employs temporal 1D-DCT on inter-frame residual paths in vision-latent space, capitalizing on the notable low-frequency concentration observed. To synchronize frequency-domain dynamics with inherent visual embeddings, Fre-Res incorporates a Spatial-Guided Absorber that integrates temporal residual data into the corresponding spatial anchor tokens. The approach demonstrates strong performance on both fine-grained short-video and long-video reasoning benchmarks. The paper can be found on arXiv with ID 2605.16366.