← News

ResearchSources: arXiv · 19 Nov 2026

SixthView benchmark tests how AI models reason about hidden views of 3D assemblies

Researchers introduce SixthView, a benchmark where models must predict the unseen sixth orthographic view of a 3D assembly from five observed views, testing symbolic and coordinate reasoning.

In plain words

When you see five sides of an object built from parts, the sixth side is often hidden. SixthView is a test set that checks whether AI models can work out that hidden side using logic about how parts fit together. The results show models do much better when given exact written geometry than when relying on images alone.

What was announced

The benchmark, described in an arXiv paper by researchers with a code release on GitHub, contains 602 synthetic polycube assemblies of 2 to 10 parts. Models must produce a dense grid of part labels for the missing view, which allows scoring both by exact-view match and cell by cell. The authors verified for each item that every legal assembly consistent with the given component geometry and observed views yields the same target view, so the answer is unambiguous. Image-only variants add shape recovery from rendered views, where occlusion can make different 3D shapes look identical.

The authors evaluated three vision-language models under four input conditions. Across 552 items with complete evaluations, exact-grid accuracy with images plus exact text ranged from 33.2% to 85.4%. Adding exact text to separate images improved accuracy by 19.7 to 63.0 percentage points, while adding images to exact text changed results by only -0.5 to +2.1 points. Separate image panels outperformed a merged sheet by 6.7 to 11.1 points. The authors conclude that performance depends strongly on access to exact symbolic information, with little further gain from images once geometry is specified.

What it does not show

  • The evidence is the paper abstract only; full methodology, model identities and detailed results are not covered in the supplied passages.
  • All assemblies are synthetic polycube constructions, so results may not transfer to real-world 3D objects.
  • The three evaluated models are not named in the supplied text, and no independent replication of the results is reported.

For security and IT teams

Scope
The benchmark code is available at github.com/facebookresearch/sixthview_benchmark, per the paper's abstract.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.