LVSum: A Benchmark for Timestamp-Aware Long Video Summarization

Apple researchers have released LVSum, a new human-annotated benchmark designed to evaluate how multimodal large language models summarize long-form video content. The project addresses the difficulty models face in maintaining temporal accuracy and aligning descriptions with specific timestamps over extended durations. By providing a standardized testing ground, this benchmark aims to improve the ability of AI systems to generate summaries that are both semantically coherent and chronologically grounded.