SSoN benchmark exposes a large gap in AI spatial reasoning
Princeton AI Lab introduced Same Space or Not (SSoN), a benchmark that asks models to identify which of four static indoor views contradicts the others and explain the evidence. In results reported by 36氪 and AIbase, humans scored 83.3% on 1,000 questions, while the best of six tested VLMs, GPT-6 Astra Pro, reached 39.7%. Targeted reasoning-grounding supervision improved Qwen2.5-VL-7B from 21.89% to 41.79% on 201 questions from unseen rooms.
SSoN isolates cross-view spatial consistency from dynamic interaction and action control, revealing a substantial human–VLM gap while showing that geometry cues and reasoning-grounded supervision can improve performance.
AI 真的能理解三维世界吗？
规划下一步动作、在杯子被挪走后及时调整、赶在环境变化前完成操作，动态任务不断向 AI 提出更复杂的空间要求。支撑这些行为的一块基础，是对眼前空间的可靠理解：物体在哪里，不同视角的观察能否彼此一致？
普林斯顿 AI Lab 的研究者们带来了一张空间智能新考卷：Same Space or Not（SSoN）。它拿掉了实时变化和动作控制，只给四张静态室内图像，让模型找出其中违背同一场景的那一张，并解释为什么。
- 项目主页：https://sson.aisexam.com/
- 论文链接：https://arxiv.org/abs/2610.23959
- 数据集：https://huggingface.co/datasets/HuiziCao/SSoN
- 代码链接：https://github.com/sson-benchmark/SSoN
- 微调权重：https://huggingface.co/HuiziCao/SSoN-Qwen2.5-VL-7B-LoRA