Original title: WorldReward: Reward Modeling for Camera-Conditioned World Models
Yibin Wang, Zehan Wang, Junshu Tang, Zhimin Li +12 · September 2026
Video-generating artificial intelligence can now act as a world model. This means it creates interactive scenes based on camera movements you choose. The problem is judging if the video looks good while also following those camera actions correctly. Existing evaluation tools only check visual quality or path accuracy, not both. To solve this, researchers built WorldReward, a reward model that compares two videos at once. Instead of judging a whole video at once, WorldReward breaks paired videos into action-aligned chunksVideo segments that correspond to specific actions or commands.. It gathers structured visual evidence for each small piece, makes chunk-level decisions, and uses voting to determine video-level preferences for action and visual quality. The researchers found that WorldReward beat GPT-5.5 on a human-annotated test set across action consistency, appearance quality, and motion quality. When used to train an interactive video model called HY-WorldPlay 1.5, it consistently improved action execution and visual quality over short- to long-term horizons.
You are judging a dance performance. You need to see if the dancer follows the choreography and if the moves look graceful. Watching the entire performance at once might make you miss small details. In this paper, the model divides the video into segments based on each action, inspects each segment for both action accuracy and visual appeal, and then combines those checks to decide which video is better overall.
Camera-conditioned world models generate interactive videos, but evaluating them separately for actions and visuals misses their connection. WorldReward provides a unified assessment, helping developers build more coherent and realistic videos for applications like simulations and games.
The abstract doesn't state limitations, but it is specific to camera-conditioned world models and depends on a vision-language model.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
WorldReward: Reward Modeling for Camera-Conditioned World Models — Yibin Wang, Zehan Wang, Junshu Tang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.