Researchers from Shanda AI have introduced WildWorld, a massive dataset designed to bridge the gap between simple video generation and complex world modeling. By leveraging high-fidelity footage from the AAA game Monster Hunter: Wilds, the team provides a sandbox for AI to learn how specific actions—like swinging a sword or dodging—affect the environment and character states. Unlike previous datasets that focus solely on pixels, WildWorld includes detailed metadata like character skeletons and camera poses.

This state-aware approach addresses a major hurdle in generative AI: long-horizon consistency. Current video models often suffer from drift, where the scene becomes nonsensical over time because the AI doesn't understand the underlying rules of the world. By training on explicit state transitions rather than just visual changes, models can better maintain logical flow during extended sequences. This is a critical step toward creating generative Action Role-Playing Games (ARPGs) where the world reacts dynamically to player input.

The release includes WildBench, a benchmark designed to test how well models follow complex action prompts. Early results suggest that even advanced models still struggle with semantically rich actions, highlighting a significant frontier for future research. This dataset provides the structured data necessary to move AI from mere video mimics to systems that understand the physics and logic of the digital worlds they inhabit.

Researchers from Shanda AI have introduced WildWorld, a massive dataset designed to bridge the gap between simple video generation and complex world modeling. By leveraging high-fidelity footage from the AAA game Monster Hunter: Wilds, the team provides a sandbox for AI to learn how specific actions—like swinging a sword or dodging—affect the environment and character states. Unlike previous datasets that focus solely on pixels, WildWorld includes detailed metadata like character skeletons and camera poses.

This state-aware approach addresses a major hurdle in generative AI: long-horizon consistency. Current video models often suffer from drift, where the scene becomes nonsensical over time because the AI doesn't understand the underlying rules of the world. By training on explicit state transitions rather than just visual changes, models can better maintain logical flow during extended sequences. This is a critical step toward creating generative Action Role-Playing Games (ARPGs) where the world reacts dynamically to player input.

The release includes WildBench, a benchmark designed to test how well models follow complex action prompts. Early results suggest that even advanced models still struggle with semantically rich actions, highlighting a significant frontier for future research. This dataset provides the structured data necessary to move AI from mere video mimics to systems that understand the physics and logic of the digital worlds they inhabit.