VideoMM is introduced, which marks a paradigm shift from model-centric downsizing to adaptive perceptual granularity and accelerates inference by 2.73 times over current leading methods, establishing a highly scalable paradigm for long-video understanding.
Abstract
Scaling Multimodal Large Language Models (MLLMs) to long-form video understanding is bottlenecked by the explosion of visual tokens, which saturates context windows and incurs prohibitive costs. Current solutions predominantly rely on auxiliary models for token reduction but face a fundamental dilemma: lightweight encoder-driven approaches often overlook critical semantic information, whereas heavyweight MLLM-driven reduction negates the efficiency gains. {In this work, we identify a more fundamental inefficiency underlying this dilemma: while fine-grained visual details are essential for detailed understanding, they are largely redundant for the preliminary task of selecting semantically relevant regions. } Motivated by this, we introduce \textbf{VideoMM}, which marks a paradigm shift from model-centric downsizing to adaptive perceptual granularity. Specifically, our framework {decouples selection from reasoning} by executing semantic filtering on a cost-effective \textit{Macro Proxy} (derived from downscaled frames), and projecting the selected regions onto high-fidelity \textit{Micro Tokens} for detailed understanding only when necessary. Extensive evaluations show that VideoMM significantly outperforms existing solutions. It achieves a 6.13$\times$ speedup and a 7.4\% accuracy gain over full-context baselines on LongVideoBench, and further accelerates inference by 2.73$\times$ over current leading methods, establishing a highly scalable paradigm for long-video understanding. Our code is available at: https://github.com/adfh917k/VideoMM.
The key design insight is hierarchical ordering: by first compressing spatial dimensions, VETO drastically reduces the cost of subsequent global temporal matching, bypassing the efficiency wall of single-axis approaches, with an advantage that grows with modern fully-fused attention infrastructure.
PACE (Pixel-Adaptive Condense and Extract), a training-free inference framework that accelerates both the vision encoder and the Large Language Model (LLM) via a unified Condense-and-Extract paradigm, is proposed.
KeyRec is a training-free framework for constructing bounded visual memory that preserves fine-grained recent observations in a visual cache and organizes historical evidence into a structured event bank and achieves the best compressed performance in 13 of 15 settings using only 10\% of the dense decoder-facing visual...
Zi-Han Chen, Xuejian Rong, Xiao-Juan Wang et al.· 0 citations
Video Large Language Models (VideoLLMs) have achieved strong video understanding capabilities but incur substantial inference overhead due to the large number of visual tokens. Existing VideoLLM token compression methods largely rely on selection-independent scoring, overlooking cross-frame complementarity and conseque...
Shuo Yang, Chang-Bai Li, Rui Tang et al.· 0 citations
Vision-language models (VLMs) have demonstrated impressive capabilities but suffer from substantial computational overhead, as vision tokens dominate the input sequence. This motivates vision token compression as a key direction to alleviate the burden. However, with the emergence of hybrid architectures incorporating...
Hao Jiang, Yi-Ru Mao, Tian-Peng Bu et al.· Proceedings of the 2026 Conf...· 0 citations
Multimodal large language models (MLLMs) have recently extended from static image understanding to video comprehension, but representing videos as frame-level token sequences incurs substantial computational overhead. Existing visual token compression methods typically rely on uniform sampling or single-dimension redun...
Jian-Xin Ma, Shi-Bo Jin, Lu-Juan Dang et al.· IEEE Transactions on Image P...· 0 citations