VideoMM is an adaptive macro‑micro inference framework that tackles the visual token explosion in long‑form video understanding. It first downscales video frames into a Macro Proxy, performs semantic filtering on this low‑cost representation to select task‑relevant regions, and then projects only those regions onto high‑fidelity Micro Tokens for fine‑grained reasoning. This decouples selection from reasoning, preserving essential semantics while dramatically cutting computation. Experiments on LongVideoBench show that VideoMM is 6.13× faster than full‑context baselines with a 7.4% accuracy gain, and further outperforms the current state‑of‑the‑art by 2.73×. The code is publicly released.
Review: By filtering semantics at a coarse level before applying detailed analysis, VideoMM delivers a scalable balance of speed and accuracy, setting a new benchmark for long‑video comprehension.