TuringSpace tests spatial intelligence through interactive CAPTCHAs
Xuwei Ding and 11 coauthors introduced TuringSpace, a synthetic web environment containing 125 interactive CAPTCHA tasks across five spatial abilities. The paper reports that humans solved 94.1% of the tasks, while GPT-6 Astra failed 25.9% and Kimi K3 failed 68.0%. Fine-tuning Qwen3.5-9B with the environment raised its success rate from 1.6% to 33.1% and produced gains across seven out-of-distribution benchmarks.
The preprint presents an interactive benchmark and training environment for evaluating whether multimodal models can perceive, act, and correct spatial decisions, with reported gains transferring to real-world scenes.
Computer Science > Machine Learning
arXiv:2611.25387 (cs)
[Submitted on 19 Nov 2026]
Title:TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs
Authors:Xuwei Ding, Skylar Zhai, Linxin Song, Jiarui Feng, Dongqi Fu, Huajun Xi, Junnan Li, Jingcheng Liang, Jieyu Zhao, Yinglong Xia, Hong Li, Jian Kang
View a PDF of the paper titled TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs, by Xuwei Ding and 11 other authors