LLM Spatial Reasoning Limited in Complex Game Tasks

A new study on arXiv (2607.22732) investigates spatial reasoning failures in LLM-based game agents, testing whether causal prompt augmentation and multi-step planning improve win-rates. Using the Qwen3 model family across varying scales, reasoning modes, and planning horizons, researchers introduced a GVGAI benchmark with three custom games at five difficulty levels. Experiments included a positioning experiment for coordinate accuracy and gameplay success evaluation. Results show larger models with thinking mode identify positions more accurately, but smaller models struggle with coordinate matching. Win rates decline with increasing game complexity.