Autoregressive Mosaics introduces AM-Bench: a benchmark that separates translating given geometry into code from composing a layout from a text prompt. We evaluate eight open-weight models (8B–34B) on 150 layout prompts, producing 13,200 generation attempts.
How well can text-only language models compose 2D layouts, how does the output medium affect this ability, and what spatial information do they represent before drawing?
Geometry is easier than composition. All eight models reach a median part-wise intersection-over-union of 1.0 when geometry is specified. Their ability to compose a scene varies substantially.
The medium matters. Switching from the custom Python drawing API to SVG improves pooled layout scores by 0.37 points on a 0–5 scale, at the same 24 × 24 resolution.
A coarse layout is decodable before drawing. Probes outperform a text baseline in all eight models. Further analysis of two models recovers shared layout structure, but not their specific output layouts—consistent with incremental construction rather than a complete advance plan.
These results concern coarse, code-generated images; layout quality is assessed by two vision-language judges.
An early prototype of generating mosaics through text and code.
Exploratory ASCII and code-canvas outputs from the prototype, separate from the benchmark evaluation.
Models write Python using six drawing primitives on a deterministic 24 × 24 canvas. Two tasks separate instruction translation from spatial composition.
Translate fully specified shapes, positions, and colors into drawing code. Part-wise intersection-over-union measures agreement with exact reference geometry.
Specified LayoutExact ReferencesChoose geometry and placement from an underspecified prompt. Two vision-language judges score fidelity, shape, color, spatial arrangement, and completeness across elemental, iconic, and compositional prompts.
150 PromptsThree Tiers@misc{nedungadi2026autoregressivemosaics, title = {Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models}, author = {Nedungadi, Ashwin and Oehmcke, Stefan and Lüdtke, Stefan}, year = {2026}, eprint = {2608.30751}, archivePrefix = {arXiv}, primaryClass = {cs.AI}, url = {https://arxiv.org/abs/2608.30751} }