WorldMemBench：视频世界模型记忆能力综合基准
原标题：WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
AI 摘要
研究者提出 WorldMemBench，一个用于系统评估视频世界模型「世界记忆」能力的基准，包含 260 个测试用例（200 个静态记忆、60 个动态记忆），覆盖人类身份与外观、物体外观与几何、环境内容与几何的保持，以及事件推进与状态持续性。评测九款代表性模型发现，强静态保持能力并不代表可靠的动态状态演化，重复回访会暴露记忆退化。人类研究支持自动指标与人类判断的一致性。
正文节选
WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models Abstract As video models enable longer and more realistic interactive experiences, maintaining a world consistent with its history becomes essential. This requires both preserving previously observed content and ensuring coherent world evolution over time. World memory is a critical capability to evaluate alongside visual fidelity and video duration. We introduce WorldMemBench, a benchmark for systematically evaluating wor