Apple-π benchmarks physical reasoning in video AI

While current video generation models produce aesthetically impressive clips, they frequently rely on visual heuristics rather than real physical principles; Apple-π provides a much-needed diagnostic framework for true law-grounded physical intelligence.

• Evaluates models across three distinct reasoning stages—perception, formulation, and deduction—to pinpoint exact failure modes.

• Includes the Orchard dataset featuring 400 physics video scenarios across ten fundamental classical mechanics tasks.

• Establishes a benchmark requirement for world models to demonstrate grounded scientific reasoning instead of shortcutting visual output.

• Provides critical benchmarking infrastructure needed for trustworthy embodied AI and physical world simulation.