The Sword, Shield, and Achilles' Heel: Characterizing the Linguistic Inductive Bias of Large Language Models for Spatial Reasoning in Navigation Planning

Researchers tested how linguistic format and context shape LLM navigation planning using a dual-interventional framework.

Across multiple spatial reasoning benchmarks, topological information consistently improved robustness and planning.

Effects from compression and wording varied by model and task, showing strong linguistic inductive bias.

Incorrect semantic cues could seriously mislead models, causing navigation failures.