Reddit user creates maze benchmark for LLM spatial reasoning

The benchmark requires models to navigate a maze, find a key, and unlock an escape door to test spatial awareness and memory. The creator found notable differences in performance across models, with some failing at simple pathfinding.