Home / AI Research / TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs
TUAI Research
TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs
by meta · Xuwei Ding, Skylar Zhai, Linxin Song, Jiarui Feng, Dongqi Fu, Huajun Xi
Humans act in the visual world by inferring the 3D layout behind a 2D view, predicting how an action changes it, and correcting themselves after seeing the result. Can multimodal large language models (MLLMs) do the same? CAPTCHAs offer a natural test, since humans solve them through this very loop, clicking and dragging on the screen and adjusting after each result. Yet existing CAPTCHA benchmarks are largely solved and rarely spatial, while spatial benchmarks rarely require action. We therefore present TuringSpace, a synthetic web environment of 125 interactive CAPTCHAs that span five spatial abilities and must be solved through clicks and drags on the screen. Every task is a program…
Upvotes23
Comments2
Published18 Nov 2026
More from meta
| MUMuse EmberAI Model APIs · #17 | $0.10 / $0.50 |
| MUMuse Spark OpenAI Model APIs · #39 | $1.25 / $4.25 |
| MUMuse WonderAI Model APIs · #48 | $4.00 / $20.00 |
| MUMuse Spark 1.3 ContributorAI Model APIs · #97 | $0.10 / $0.20 |
| MUMuse Spark 1.3AI Model APIs · #98 | $1.25 / $4.25 |
| MUMuse Spark 1.2 ContributorAI Model APIs · #106 | $0.10 / $0.20 |