LLMs struggle with sequential spatial reasoning: prediction accuracy drops sharply as operation histories grow longer, and models with similar overall performance can differ dramatically in which specific cells they get wrong.
Grid-Bench is a benchmark that tests whether large language models can predict how discrete worlds change when multiple spatial operations are applied in sequence. Models read initial grid states, rules, and operation histories as text, then predict final grids—with reference execution enabling precise measurement of prediction accuracy across different spatial dimensions and operation types.