Benchmark RadarVideoRover-Bench
VideoRover-Bench is a benchmark for open-world video reasoning that combines video understanding with deep research, stratified by video duration and research difficulty. It evaluates the capability of models to locate sparse visual evidence and acquire external knowledge through coordinated video cropping, multimodal search, and webpage browsing.
- Released
- 2026-08-24
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
The benchmark fills a gap by assessing unified video reasoning and multi-step information seeking, which are typically developed in isolation. It provides a structured evaluation to guide the development of video agents that can handle complex open-world tasks requiring external knowledge.
Motivation
Open-world video understanding often requires a model to locate sparse visual evidence and acquire external knowledge that is absent from the video and its parametric memory.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.