[id: 105]

Geospatial Algorithm Discovery driven by LLMs: Spatial Regime Model as an Example

Short description: The aim of this thesis is to investigate how an LLM-driven workflow for geospatial algorithm discovery, using the Spatial Regime Model as an example, can support the generation, evaluation, and understanding of spatial analytical methods.

Keywords:  Large Language Models, Algorithm Discovery, Geospatial Models, Domain Knowledge

Topic at: TU Munich

Staff involved: Xiayin Lou (xiayin.lou@tum.de)  

Description:

In recent years, large language models (LLMs) have shown their great power in helping people understand and solve complex problems. It is possible to accelerate and automate the knowledge discovery procedure by employing the established LLMs, such as GPT-5 and Gemini 3.0 (Novikov, et. al., 2025). Combining the evolutionary process with LLM-based code generation and geospatial knowledge, GeoEvolve (Luo, et. al., 2025) can automatically explore novel geospatial models (see figure above). However, the evolved geospatial algorithms still need evaluation in terms of effectiveness, efficiency, code quality and generalization, limitations.

The research will:

- Build an LLM-driven algorithm discovery workflow using GeoEvolve.

- Apply the workflow to generate and implement a Spatial Regime Model.

- Evaluate the generated algorithm in four key dimensions

1) Effectiveness: model correctness, stability, analytical quality

2) Efficiency: runtime, resource usage, scalability

3) Code Quality: readability, maintainability, structural correctness

4) Generalization: performance across different datasets and spatial contexts

- Additionally, the study will analyse limitations of LLM-generated algorithms.

- Derive insights from GeoEvolve about the nature of algorithm discovery in geospatial tasks.

The figure below visualizes the Workflow of Geospatial Algorithm Discovery

Literature/references:

  1. Luo P, Lou X, Zheng Y, Zheng Z, Ermon S. GeoEvolve: Automating Geospatial Model Discovery via Multi-Agent Large Language Models. arXiv preprint arXiv:2509.21593. 2025 Sep 25.
  2. Novikov A, Vũ N, Eisenberger M, Dupont E, Huang PS, Wagner AZ, Shirobokov S, Kozlovskii B, Ruiz FJ, Mehrabian A, Kumar MP. AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131. 2025 Jun 16.