LLMs Enhance Spatial Reasoning with Building Blocks and Planning

This paper addresses the challenge of spatial reasoning in LLMs, a crucial capability for applications like navigation and planning. The authors propose a novel two-stage approach that decomposes spatial reasoning into fundamental building blocks and their composition. This method, leveraging supervised fine-tuning and reinforcement learning, demonstrates improved performance over baseline models in puzzle-based environments. The use of a synthesized ASCII-art dataset and environment is also noteworthy.

• •Proposes a two-stage approach for spatial reasoning in LLMs.
• •Uses supervised fine-tuning for elementary spatial transformations.
• •Employs reinforcement learning with LoRA adapters for multi-step planning.
• •Outperforms baselines in puzzle-based environments.
• •Utilizes a synthesized ASCII-art dataset and environment.