Benchmark RadarAI BENCHMARK PROFILE
MBench
MBench is a benchmark for memory capability of video world models, decomposing into entity, environment, and causal consistency with 12 sub-dimensions. Includes code, dataset, and leaderboard.
- Released
- 2026-05-30
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Fills the gap in evaluating long-term state retention in video world models, providing a standardized benchmark to advance the field.
Motivation
Recent advancements in video-based world models have demonstrated an unprecedented ability to synthesize high-fidelity visual sequences.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.