What does your world model need to prove?
Updated 2026-10-12
Choose an evaluation by the question it answers: following a camera path, responding to actions, respecting physics, supporting a robot or reasoning about space. Entries explain the tested scope and where it stops.
This is a directory of protocols, not a combined leaderboard. Compare models within the same benchmark version and settings; follow the primary project for code, data and current results.
Find a measurement protocol
Browse six research questions with source-linked scope notes. Video generators, interactive worlds and adjacent agent tests are distinguished.
Read scores without false comparisons
Check versions, test settings and task coverage before using a reported result. Scores from separate suites are not interchangeable.
Understand the evaluation terms
Plain-language definitions of rollout, non-reactive evaluation, action conditioning and related concepts.
Submit a benchmark or source update
Maintainers can suggest reproducible resources, report an error or propose a focused collaboration through one form.
Examples of different measurement scopes
- WorldScore measures camera-conditioned world generation across control, quality and dynamics. [haoyi-duan.github.io]
- Physics-IQ tests physical behavior in generated video across five physics categories. [physics-iq.github.io]
- NAVSIM evaluates driving policies with a non-reactive environment, a different task from generated-video quality. [arxiv.org]