Researchers have developed a new method for analyzing hours-long egocentric video footage using vision-language models. Current methods are expensive because they require decoding the entire video and making multiple model calls. The proposed 'audio-first triage' approach selects which parts of the video to analyze based on audio cues, reducing the number of model calls needed while maintaining coverage of important actions. This method uses frozen AudioSet-pretrained features without requiring domain-specific sound-event labels. The results show that it outperforms previous methods in terms of action coverage and reduces the number of model calls by 9-20%. --- Why it matters: This matters to engineers working on AI-powered video analysis because it provides a more efficient way to process large amounts of egocentric video data, which is essential for applications such as progress monitoring, quality control, and safety in various industries. By reducing the number of model calls needed, this approach can help make vision-language models more practical for real-world use. Source: https://arxiv.org/abs/2608.22359