LLM-guided path planning

Overview

My team project for CSCE 625 Artificial Intelligence at Texas A&M, running through Fall 2026. We're testing whether a large language model can guide heuristic search through generated maps more efficiently than hand-written heuristics.

The problem

Heuristic search is only as good as its heuristic. Simple estimates like straight-line distance ignore a map's structure, so search can waste effort exploring dead ends. Language models can reason about a map as a whole, but it's an open question whether their guidance actually makes search cheaper once you count the cost of asking them.

What I did

1. 1.Built the environment and map generator that every experiment runs on.
2. 2.Integrated the LLM into the search loop and wrote the functions that evaluate the guidance it gives.
3. 3.Worked alongside my teammate, who owns the search algorithms, benchmarking and visualizations.
4. 4.Extending the setup to pursuit-evasion, where minimax search plans for both the pursuer and the evader.

Results

• Results land at the end of the semester in December 2026. I'll update this page with the numbers.