Graph-Enhanced Large Language Models for Spatial Search

TL;DR

A new framework is proposed to enhance LLMs' spatial reasoning abilities by leveraging graph data.

Problem

While LLMs have improved in complex tasks and domain-specific Q&A, their spatial reasoning abilities, crucial for physical-world domains like urban planning, civil engineering, and travel, are still lacking.

Approach

Envisions integrating search engines with LLMs to answer complex spatial questions through graph-enhanced reasoning, noting that spatial data is commonly stored as graphs. Analyzes challenges in spatial reasoning and proposes directions for graph-based LLM development.

Results & Contribution

Systematically outlines challenges in spatial reasoning and emphasizes the need for new research techniques combining graph data with LLMs to enable impact in relevant domains.