InfiniMind and University of Tokyo Release NARU Bench for Japanese Video AI
NARU Bench directly targets the gap between current video AI benchmarks, which are English-centric and short-form, and the demands of real Japanese media where implicit social cues and long-form narrative coherence are essential.
Reporting from 1 source: ASCII.jp.
InfiniMind Inc., in collaboration with the University of Tokyo, has released NARU Bench, a benchmark designed to evaluate AI's ability to understand high-context Japanese video content. The benchmark includes 1,481 questions from 155 YouTube videos totaling 147 hours, testing narrative tracking and cultural reasoning. It is available on Hugging Face and GitHub.
The benchmark was developed under the government-led GENIAC initiative. InfiniMind and the University of Tokyo designed it to measure what they call a 'new evaluation axis' for video understanding, focusing on abilities like reading the air and interpreting true intentions versus public facade. The dataset draws from 155 Japanese YouTube videos, including interviews and talk shows ranging from 30 minutes to four hours. Questions are written in Japanese and require watching the video to answer, not relying on transcripts or prior knowledge. NARU Bench covers two tracks: Narrative Understanding with 745 questions and Cultural Understanding with 736 questions.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.