Researchers test whether large language models can navigate unknown environments using only text commands without code execution or external tools.

Study evaluates nine contemporary LLMs (both open-source and proprietary) on text-based navigation tasks in ASCII gridworlds with only partial observability

LLMs must control agents using only UP/RIGHT/DOWN/LEFT commands while viewing only a 5×5 local window at each step

Two primary tasks tested: Exploration (maximizing revealed cells) and Navigation (reaching goals via shortest paths) across three difficulty levels

Research examines whether reasoning-tuned and instruction-tuned models of different architectures (dense and Mixture of Experts) can function as effective text-only controllers