Meta (Xuwei Ding et al.) · Spatial Action Benchmark

TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs

TuringSpace is a synthetic web environment of 125 interactive CAPTCHAs covering five spatial abilities, solved through clicks and drags, with each task generating new instances, a checker and an oracle. Humans solve 94.1% of tasks, the best model GPT-6 Astra fails 25.9%, and the best open-weight model Kimi K3 fails 68.0%. Fine-tuning Qwen3.5-9B on the gym raises its success rate from 1.6% to 33.1%, with gains across seven out-of-distribution benchmarks, including 9.9 points on VSI-Bench indoor videos. The authors state they will release TuringSpace.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.