Back to articles

TuringSpace Tests Spatial Intelligence Through Interactive CAPTCHAs

3 min read

Introduction

Recognizing objects in an image is not the same as understanding how they are arranged in space. Acting in a visual environment requires a loop: infer the three-dimensional scene from a two-dimensional view, predict what an action will change, then inspect the outcome and adjust. TuringSpace makes that loop the subject of evaluation. Rather than choosing an answer from a list, a model must operate a browser using clicks and drags.

Key findings

  • A benchmark built around action. TuringSpace is a synthetic web environment with 125 interactive CAPTCHAs spanning five spatial abilities. Tasks are solved through on-screen clicks and drags, and agents must use the resulting visual feedback to decide what to do next.
  • A substantial human-model gap. Humans solve 94.1% of the tasks. GPT-6 Astra, the best-performing model in the reported comparison, reaches 74.1%, leaving 25.9% unsolved. The best open-weight model, Kimi K3, fails 68.0%. The results underline that visual recognition alone does not ensure reliable action in a visual scene.
  • Perception is a recurring failure point. The authors report that most failures stem from misreading screenshots. Even when an action produces feedback, model reasoning rarely corrects a mistaken belief; instead, it may generate evidence that supports the original interpretation. More reasoning, by itself, may therefore be insufficient when the initial visual reading is wrong.
  • The benchmark doubles as a training environment. Each task is generated by a program and comes with a checker and an oracle solution, allowing fresh instances and verified training trajectories. Fine-tuning Qwen3.5-9B on 55K trajectories lifts its success rate from 1.6% to 33.1%, above Kimi K3. The tuned model also improves on all seven reported out-of-distribution benchmarks, including a 9.9-point gain on real-world indoor videos in VSI-Bench.

Why it matters—and what it does not show

TuringSpace’s contribution is not just another leaderboard. It frames spatial understanding as a sequence of perception, action, and correction, making it harder for a system to succeed through image recognition or verbal reasoning alone. Click-and-drag tasks can reveal whether a model understands spatial relationships and anticipates how an action changes a scene. Programmatic task generation and verified trajectories also make the benchmark useful for training, not only evaluation.

At the same time, even the fine-tuned model remains well below human performance on TuringSpace. The reported gains suggest that these capabilities can be improved with targeted data, but do not establish general spatial intelligence. Further testing is needed to determine whether models can reliably perceive and revise their beliefs across a wider range of real-world settings, including computer-use and robotic pointing tasks.

Source: Hugging Face Daily Papers

Comments

Checking sign-in status...

Loading comments...

Related articles