Advancing Spatial Reasoning in Large Language Models: An In-Depth Evaluation Using the StepGame Benchmark

Our summary

Evaluates LLMs on the StepGame spatial-reasoning benchmark, finding they map language to spatial relations reasonably but degrade on multi-hop spatial inference; proposes prompting and neuro-symbolic enhancements.

Why it matters

Pins spatial and geometric reasoning failures to multi-hop composition over relations — a concrete, reproducible weak spot.

Related findings (1)

• Spatial and geometric reasoning errorsLow

  Models struggle with relative positions, rotations, and simple geometric/visual reasoning.

Cite this