HappyWorld-Bench Tests Whether World Models Stay Reliable in Interaction

Introduction

World models are increasingly judged by more than the realism of their outputs. A system may generate an attractive video or a convincing 3D scene, yet fail when an agent explores the environment, performs several actions, returns to an earlier location, or changes part of the world. HappyWorld-Bench is designed to measure this less visible but more important property: whether a generated world remains dependable under interaction and intervention.

What the benchmark covers

• Three evaluation tracks. The benchmark separately examines video world models, spatial world models, and embodied world models, while applying a common reliability-oriented perspective.
• A six-level capability framework. World capabilities are organized into six stages, W1 through W6, progressing from generative construction toward unified world modeling.
• Behavior rather than appearance. Tests target exploration, extended rollouts, revisits, object placement, scene edits, multi-step actions, and responses to changed physical or action conditions.
• Human and automated evaluation. HappyWorld-Arena runs human A/B comparisons and converts them into model-level Elo ratings. Newly designed automated metrics measure whether a system’s behavior is correct, providing a complement to preference judgments.
• A broad test set. The benchmark includes 1,138 video prompts, 300 spatial scenes, and 254 embodied cases. It evaluates 14 video models, 9 spatial systems, and 8 embodied candidates.

Main findings

The results point to reliability problems across all three tracks. Video models tend to lose consistency during extended rollouts and when previously visited content is revisited. A short clip can look coherent without the underlying world remaining stable over time. This distinction matters for agents that must plan across multiple observations rather than consume a single generated sequence.

Spatial systems also show a gap between understanding an instruction and carrying it out precisely. The best reported placement accuracy is 70.14%, while edit execution reaches 73.33%. These figures indicate that models can handle part of the requested spatial behavior, but still make substantial errors when objects must be positioned or scenes must be changed according to instructions.

Embodied candidates struggle to preserve state across multi-step actions. They also have difficulty responding precisely when action conditions or physical rules are altered. In an embodied setting, such failures can propagate: an incorrect state estimate at one step can make later planning and control unreliable.

Why it matters

HappyWorld-Bench shifts the evaluation question from “Does the generated world look realistic?” to “Can an agent continue to trust and use this world?” Its design separates visual quality from consistency, action response, and behavior after intervention, then brings those dimensions together in one evaluation framework.

For developers, the findings suggest that better world models will need more than improved rendering or generation. They will require stronger state tracking, longer-horizon consistency, action-conditioned prediction, and robustness to rule changes. For users of simulation, games, robotics, and agent systems, the benchmark offers a way to distinguish impressive demonstrations from systems that can support sustained interaction. The central challenge is not producing a plausible world once, but preserving its logic as the world is explored and changed.

Source: Hugging Face Daily Papers