Grid-Bench Evaluates LLM Spatial Reasoning via Discrete Grid Transformations

Grid-Bench measures LLM accuracy in predicting state changes across 2D, 3D, and 4D grids following sequential spatial operations. Results show that complete-state accuracy decreases sharply as the history of operations grows, even among models with similar single-event accuracy.