Machine Heart

Sep 15, 2026 · Artificial Intelligence

WorldRoamBench: Benchmarking Long-Horizon Stability in Interactive World Models

Amap, Nanjing University, Tsinghua, and Peking University release WorldRoamBench, a benchmark with 1000+ long-horizon roaming samples evaluating 12 interactive world models on action following, visual stability, physics adherence, and memory retention, revealing that even top models like Genie 3 show significant weaknesses in specific dimensions.