LLM Linguistic Bias in Spatial Navigation Planning

A new study from arXiv introduces a dual-interventional framework to evaluate how linguistic structures in text-based spatial representations affect Large Language Model (LLM) performance in navigation planning. The framework separates representation intervention, which varies linguistic format and compression, from context intervention, which combines contextual features like topology and geometry. The research aims to characterize LLMs' linguistic inductive bias, challenging the assumption that textual encoding of spatial data is a neutral engineering choice. The paper is available as arXiv:2605.31404v1.