浙大提出 ProVisE 框架，让图像生成模型直接画出空间认知答案
用视觉协议让生成模型以画图方式回答空间问题，替代传统的坐标输出，更贴近人类直觉。
现有空间认知评测强迫 AI 用坐标或文字作答，这对图像生成模型并不自然。浙大团队提出 ProVisE，用视觉协议引导模型直接在图上标记、画轨迹或生成深度图，再自动解析为可评分的结果，并构建了覆盖 14 项子任务的 SpatialGen-Bench。
正文摘录
.jpg)  If you ask someone "Where is the cup?", they'd normally just point at it, rather than replying robotically: "The target is located within a bounding box spanning x-coordinates 345–412 and y-coordinates 512–600." Yet in existing spatial cognition benchmarks, we've been forcing AI to prove its spatial abilities through exactly this kind of counterintuitive coordinate-based output. ![图片](https://mmbiz.qpic.cn/szmmbizpng/5L8bhP5dIqHub3FiaNlBQu0oe6AyC3JwMYXAUYlvJCkzSCGOgqnKTUeYpFEGs8kqSo2pGchf1FDtXaJfj9ub4s…