Long video understanding requires reasoning over hours of content that far exceed the context window of multimodal large language models (MLLMs). According to a new arXiv paper, existing approaches attempt to handle this by compressing video into fixed-length clips, but this fragments events and undermines coherent understanding.
The paper introduces EMMA, described as an event-driven multimodal memory system with multi-agent retrieval. The abstract suggests this design is intended to address the limitations of fixed-length compression, though it does not provide details on the architecture or experimental results.
As the abstract is brief and truncated, the full methodology and evaluation remain to be seen. The paper is available on arXiv under identifier 2610.23892.