NARU: New Benchmark for Narrative and Cultural Understanding in Japanese Long Videos

A new benchmark named NARU has been developed by researchers to assess narrative progression and cultural comprehension in Japanese long-form videos. This benchmark features 1,481 questions derived from 155 videos, which collectively run for 146.8 hours, covering four narrative and five cultural dimensions. To achieve this extensive benchmark, the team implemented a hierarchical memory-based annotation process that converts raw video into organized annotations related to events, narratives, and culture, followed by question generation through task-oriented synthesis and iterative shortcut elimination. The construction involved two verification stages by native speakers. This initiative fills a gap in current benchmarks, which seldom assess narrative tracking alongside implicit social meanings, especially in high-context, non-English media. The research can be found on arXiv with the identifier 2608.13210.