Researchers propose a new framework called COMET for video multimodal large language models. The current limitations of these models are addressed by introducing a temporal motion branch that represents frame-to-frame change and enables appearance-motion interaction. This is achieved through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. The method shows consistent improvements in action-centric tasks, temporal reasoning tasks, and static perception tasks on various benchmark datasets. --- Why it matters: This matters to AI researchers because it addresses a key limitation of current video multimodal large language models: their fragile fine-grained motion-temporal understanding. By improving this aspect, the model's ability to reason about temporal relationships in videos can be enhanced, which is crucial for applications such as video analysis and understanding. Source: https://arxiv.org/abs/2608.21030