Multimodal large language models (MLLMs) struggle with long-video understanding because attention scales quadratically and context windows are finite. A new arXiv preprint tackles this by focusing on frame selection: choosing which frames to feed the model rather than processing everything.
The paper, titled "From Label Priors to Task Evidence: Efficient Long-Video Frame Selection via Bayesian GRPO," proposes a Bayesian variant of GRPO (a reinforcement learning algorithm) to select frames. The abstract notes that existing retrieval-based approaches typically rely on some unspecified prior, and the new method aims to shift toward task-derived evidence.
Because the abstract is cut off, the exact mechanism and experimental results are not available from the source. The contribution appears to be a principled way to make frame selection more efficient, but readers should consult the full paper for specifics.