LLMs for Text-Based Exploration and Navigation Under Partial Observability

Researchers evaluated whether large language models can function as text-only controllers for navigation and exploration in unknown environments under partial observability. Testing nine contemporary LLMs on ASCII gridworld tasks, they found reasoning-tuned models reliably complete navigation goals but remain inefficient compared to optimal paths, with few-shot prompting reducing invalid moves and improving path efficiency.

This research addresses a fundamental question about LLM capabilities in sequential decision-making under uncertainty—a critical problem space for robotics, autonomous systems, and spatial reasoning AI. The study moves beyond typical benchmarks by testing LLMs as pure text-based controllers without code execution or tool use, creating reproducible conditions that isolate language understanding from external computation. The results reveal important constraints: while reasoning-tuned models like o1 demonstrate emergent planning abilities superior to instruction-tuned variants, they still exhibit characteristic biases (preferential UP/RIGHT actions) that cause suboptimal looping under partial observability. This suggests LLMs internalize directional priors from training data rather than learning true spatial reasoning. The finding that test-time deliberation and training methodology outperform raw parameter scaling has implications for model selection in resource-constrained applications. The authors' recommendation toward hybrid approaches combining LLMs with classical planning algorithms reflects a pragmatic acknowledgment that current language models lack the algorithmic rigor for optimal sequential decision-making. For the broader AI landscape, this work demonstrates both the promise and limitations of scaling—larger models don't automatically solve structured reasoning problems. The research indicates where LLM-based systems might succeed (high-level strategy, context interpretation) and where they fail (optimal path computation, systematic exploration), informing architecture decisions for real-world deployment in logistics, inspection, and search-and-rescue applications.

• →Reasoning-tuned LLMs reliably complete navigation tasks but remain suboptimal compared to oracle paths, indicating emergent planning without algorithmic efficiency.
• →Few-shot demonstrations significantly improve LLM navigation by reducing invalid moves and path length, suggesting prompt engineering is critical for control tasks.
• →Training methodology and test-time deliberation predict control ability better than model size, challenging the assumption that parameter scaling alone improves performance.
• →LLMs exhibit consistent directional action priors (UP/RIGHT bias) that induce looping under partial observability, revealing learned biases from training data.
• →Hybrid approaches combining LLMs with classical online planners offer a practical path to deployable navigation systems rather than pure language-based control.