Multimodal large language models (MLLMs) that handle long videos face a fundamental bottleneck: they must select a small, informative set of frames from a massive pool of redundant candidates. This selection is constrained by the model's visual-token budget and its reasoning ability, making the choice of frames critical for accurate understanding.

The paper introduces LDDR, short for Linear-DPP Dynamic-Resolution Frame Sampling, as a solution to this problem. The name suggests the method combines a linear determinantal point process with dynamic-resolution processing, though the abstract does not elaborate on the mechanics. The core idea is to improve the frame selection step so that MLLMs can work with long videos more effectively.

The abstract is brief, but it clearly identifies the challenge and the proposed direction. No experimental results or comparisons are included in the available text, so the effectiveness of LDDR remains to be detailed in the full paper.