Spatial Strategies, Not Actions: Vector-Quantized Geodesics as Tools for LLM-Driven Agents
This work investigates how large language models (LLMs) can gain spatial understanding by combining geometrical tools with LLM reasoning in grid-world environments. By using vector quantization to discover reusable movement strategies from collected shortest paths, the system achieves high goal-reaching rates comparable to full thinking modes while drastically reducing decision time from minutes to seconds.