WorldCast：多步世界状态预测的视觉生成基准
原标题：WorldCast: Exploring Intelligent Visual Generation in World State Prediction
AI 摘要
来自清华、美团 LongCat、港中文、北大、MBZUAI、上交、华科的研究者提出 WorldCast 基准，用于评估多模态生成模型在多步世界状态预测中的能力。该基准包含 8 个类别共 320 个人工设计案例，评测了 12 个代表性模型。结果显示 GPT-Image-2.5 System 得分 72.8，展现出初步的真实世界视觉状态预测能力，而 Nano Banana Pro 仅 42.8，最佳开源模型仅 20.1，表明模型在复杂上下文、纠缠引用和重推理任务上仍有明显差距。
正文节选
1]THU 2]Meituan LongCat 3]CUHK 4]PKU 5]MBZUAI 6]SJTU 7]HUST \contribution*Work done during internship at Meituan. †Project Lead. WorldCast: Exploring Intelligent Visual Generation in World State Prediction Abstract Humans interpret multimodal observations of the world and forecast how it will evolve. Recent multimodal generative systems have demonstrated similar capabilities: they can integrate multimodal context and synthesize visual outputs that depict changes in the world. To characterize thi