Benchmark RadarAI BENCHMARK PROFILE
ScratchWorld
ScratchWorld is an offline diagnostic benchmark for world models, using Scratch projects as executable worlds. It evaluates next-state prediction, long-horizon tracking, causal attribution, and counterfactual prediction with replay-verified transitions.
- Released
- 2026-06-30
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Provides a new evaluation paradigm for world models that avoids confounds like overlap with persistent state. Enables direct comparison of model capabilities on executable consequences.
Motivation
World-model evaluations often score a predicted future by overlap with a target state or observation.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.