Live intelligence

Latest AI updates

1 updates
Research

VideoACE benchmark evaluates agentic video understanding with high-frame-rate actions

VideoACE is a new benchmark for agentic video understanding that tests whether models can use interaction to improve judgments. It contains 1,000 multiple-choice questions across six categories of fine-grained action understanding, with temporal evidence annotations and screened user hints. The evaluation environment provides video inspection tools and a user simulator, letting models adjust temporal coverage, sampling density, and spatial detail and revise judgments under feedback. Five models were evaluated across single-pass observation, autonomous tool use, and tool use with user feedback.

agentic video understandingvideo benchmarksmultimodal AI