Effect of periodicity on translation performance

Investigate whether translation performance degrades on references containing periodic structures such as repeating grids and tilings, and determine whether this effect occurs across the evaluated models.

Background

The translation stress test varies four geometric difficulty axes—part count, feature size, gaps between adjacent shapes, and aspect ratio—but does not include periodicity. The reported translation task is therefore primarily a test of transcribing explicitly specified geometry, rather than generating repeated structures through loops or other compact procedures.

The authors identify periodicity as a particularly relevant unresolved case because repeating grids and tilings may invite generative looping, and because the pilot model's weakest layout subcategories included Patterns and Symmetry. The paper consequently leaves open whether the near-perfect translation results extend to periodic references, especially across the full eight-model roster.

References

The axis most likely to matter for transcription --- periodicity, the repeating grids and tilings that invite a generative loop instead of item-by-item placement --- was not built into the ladder and remains untested.

Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models  (2608.30751 - Nedungadi et al., 31 Aug 2026) in Appendix, Section 'The Translation Task (Control)', subsection 'Translation stress-test details'; elaborated in Appendix, Section 'Limitations, full detail' (Appendix~\ref{app:limitations_full})