SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence

Researchers introduced SVI-Bench, a sports-based benchmark for strategic video intelligence.

It uses broadcast basketball, soccer, and hockey data with structured annotations across nine tasks and four difficulty levels.

Models perform reasonably on perceptual questions but drop sharply on causality, simulation, and planning.

The hardest tasks require evidence gathering across millions of clips, exposing major gaps in current multimodal and agentic systems.