TuringSpace
by Xuwei Ding, Skylar Zhai, Linxin Song et al.
What is TuringSpace? TuringSpace is an agents & tool use eval made by Xuwei Ding, Skylar Zhai, Linxin Song et al.. A browser environment of 125 interactive CAPTCHA-style puzzles that test spatial reasoning through action: the model sees a full-screen screenshot, clicks and drags, looks at the result and presses Verify when it believes the puzzle is solved. Tasks are split evenly across five abilities (spatial relations, closed-loop manipulation, multi-step planning, mental rotation or folding, and tracking objects over frames). Every task is a program that generates fresh instances from a seed together with a checker, so the same tasks double as a training gym; the benchmark runs each on three fixed seeds and reports pass@1 and pass^3. The paper (arXiv:2611.25387) reports people solving about 94% of tasks while the best model tested still fails about a quarter, with most errors traced to misreading the screenshot.
Editor note
A browser environment of 125 interactive CAPTCHA-style puzzles that test spatial reasoning through action: the model sees a full-screen screenshot, clicks and drags, looks at the result and presses Verify when it believes the puzzle is solved. Tasks are split evenly across five abilities (spatial relations, closed-loop manipulation, multi-step planning, mental rotation or folding, and tracking objects over frames). Every task is a program that generates fresh instances from a seed together with a checker, so the same tasks double as a training gym; the benchmark runs each on three fixed seeds and reports pass@1 and pass^3. The paper (arXiv:2611.25387) reports people solving about 94% of tasks while the best model tested still fails about a quarter, with most errors traced to misreading the screenshot.
- What it tests
- Spatial reasoning by computer-use agents acting on interactive screen puzzles.
- Original source
- github.com/os-cua-framework/TuringSpace
- Last updated
- arXiv:2611.25387; repo last pushed 2026-11-20 (GitHub API, read 21 Nov 2026)
- License
- Apache-2.0 (code, GitHub API and README, read 21 Nov 2026)
Related agents & tool use evals
- GAIA — General assistant skills: browsing, file handling, multi-step reasoning.
- WebArena — Autonomous web navigation and task completion.
- OSWorld — Computer-use agents operating real operating systems and apps.
- Terminal-Bench — Agents working in the shell end to end.
Discussion
No one has discussed TuringSpace yet. Have you run it? Share what it caught, what it missed or how you set it up.