AM-Bench Probes 2D Spatial Reasoning in Text-Only LLMs
August 31, 2026
The Autoregressive Mosaics (AM-Bench) benchmark distinguishes between an LLM's ability to translate spatial descriptions into code and its actual internal representation of 2D geometry.
HOW THIS AFFECTS YOU
●
researcherResults indicate that while text-only models excel at geometry-to-code translation, their open-ended layout reasoning remains limited.