WROP：在世界模型中训练客体永久性
原标题：Paper page - Training Object Permanence in World Models
AI 摘要
研究者提出 WROP（World Reasoning with Object Permanence）数据基础设施，包含 150 个受认知科学启发的手工设计任务，分为六个认知类别，并通过 Blender 生成器随机化速度、光照、相机角度等参数，每个任务可生成 1 万以上样本。团队发布了 150 万样本训练语料和 300 题评测集，在评测中评估了 14 个视频模型，其中自研 16B 世界模型 PWM-WROP 在盲测成对 Elo 研究中位列续写类模型第一、总体第三。研究同时开源了数据、评测集、模型答案、分数、权重以及基于 AWS Trainium2 的原生 PyTorch 训练栈 PWM。
正文节选
Training Object Permanence in World Models Abstract Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them ideal candidates for building human-like physical intelligence. Do video models have emerged object permanence in them? If not, could we train them with a core-cognition inspired dataset? We introduce WROP (World Reasoni