FlowBench：面向空间智能的反事实想象训练基准
原标题：FlowBench: Counterfactual Imagination Training for Spatial Intelligence
AI 摘要
研究者提出反事实空间想象（CSI）任务，并发布 FlowBench 数据集与基准，包含约 42.8 万条训练样本，覆盖模拟家庭场景与真实桌面操作。实验显示，闭源视觉语言模型在需要推断运动结果时表现下降，开源模型接近随机水平；在 FlowBench-Sim-Train 上微调 Qwen3-VL-8B 后，模拟场景准确率提升 29.5 个百分点，真实图像提升 16.9 个百分点。世界模型经 FlowBench-Sim 训练后，文本到光流预测准确率也得到提升。
正文节选
FlowBench: Counterfactual Imagination Training for Spatial Intelligence Abstract Humans imagine scene changes under hypothetical motions, an ability central to spatial reasoning and robotic planning. However, this capability remains underexplored in vision-language models and world models. We formalize this capability as Counterfactual Spatial Imagination (CSI) and introduce FlowBench, a resource for training and evaluating both, with 428K training examples spanning simulated and real-world scen