The AlphaEvolve of geography? MIT and Stanford enable AI to self-evolve, understand geography, and comprehend the world.

Perhaps the daily routine of a geographical researcher is as follows: To make the required geographical model perform better, he continuously engages in conversations with large models (such as ChatGPT), attempting to improve the code or fix bugs.

The initial answers provided by large models are often imperfect. Therefore, the researcher will put forward new modification suggestions based on the results. Through such continuous interactions, the code is gradually refined.

This approach of "interacting with large models and improving algorithms" has become a common practice in geographical research.

Then the question arises: Can we take it a step further? - Can we make this interaction - improvement process happen automatically, enabling AI not just to be an assistant but to truly evolve geographical models like a scientist?

Scholars from MIT and Stanford proposed GeoEvolve, attempting such an exploration:

🌍 "Embed" geographical knowledge into AI to make its evolution more reliable and closer to geographical principles;

🤖 Transform large models from mere auxiliary tools into "research collaborators" capable of autonomously improving algorithms.

Link: https://arxiv.org/abs/2509.21593

Project address: https://vezarachan.github.io/GeoEvolveWebPage/

GeoEvolve has been open - sourced as a Python package and can be directly installed and used (pip install geoevolve).

Research Background

Geospatial modeling is a key tool for understanding climate change and promoting sustainable urban development.

However, traditional methods often rely on expert experience: formulating hypotheses, designing algorithms, and continuously adjusting parameters for improvement.

In recent years, large language models (LLMs) have demonstrated the potential for automatic code evolution. For example, Google's newly launched AlphaEvolve allows AI to attempt, mutate, and optimize algorithms on its own.

However, such systems have a natural drawback - they lack geographical knowledge. If AI is completely left to evolve on its own, it is likely to "go astray," and the generated models may lack geographical rationality.

For this reason, the proposal of GeoEvolve constructs a new framework that combines AI's autonomous evolution with geographical knowledge guidance.

GeoEvolve Framework

As shown in Figure 1, GeoEvolve can be imagined as a "GeoAI research team composed of a tutor and doctoral students":

Inner loop: AI acts as doctoral students, communicating with each other based on the initial code, constantly making trial - and - error attempts, generating, and improving algorithms;

Outer loop: A "tutor" - the geographical knowledge base (GeoKnowRAG) - provides reminders to ensure that the evolution direction conforms to spatial theory.

The four core modules of GeoEvolve are:

Code evolver (automatically generates and mutates candidate algorithms),

Code analyzer (diagnoses problems and proposes improvement ideas),

Geographical knowledge retriever (GeoKnowRAG, provides spatial theories and classic methods),

Knowledge - driven prompt generator (transforms complex geographical knowledge into optimization instructions understandable by AI).

Through this dual - loop mechanism, GeoEvolve can not only write code but also gradually learn to think like a geographer.

The entire GeoEvolve has only three inputs:

Original model - the geographical model to be improved (it can even be a blank model);

Evaluation indicators - the criteria for measuring the quality of the model (such as RMSE, MAE);

Prompt - a one - sentence task description for the large model (e.g., "Please help me improve this Kriging model").

Figure 1. The basic idea behind GeoEvolve, algorithm evolution guided by geographical knowledge. The dashed box represents a general algorithm generation engine (e.g., AlphaEvolve). The external process shows the knowledge - guided algorithm generation method proposed in this paper, specifically for geospatial modeling scenarios.

Figure 2 shows the detailed framework of GeoEvolve.

GeoEvolve aims to achieve the automatic discovery of geospatial models by combining evolutionary code generation with structured geospatial knowledge.

Different from general code agents, GeoEvolve integrates domain knowledge from spatial modeling literature and classic algorithms, enabling it to explore and discover new geospatial algorithms.

It consists of four main parts: (1) code evolver, (2) evolutionary code analyzer, (3) geographical knowledge retriever, and (4) knowledge - driven prompt generator.

These components together form a closed - loop process of code generation, evaluation, and improvement, thereby promoting the automatic discovery of geospatial models.

Figure 2. The framework of GeoEvolve

As shown in Figure 3, to prevent AI from "going astray" during the evolution process, GeoEvolve introduces a dedicated geographical knowledge retrieval module (GeoKnowRAG).

It is equivalent to a "geographical knowledge base," collecting core materials from Wikipedia, arXiv, and GitHub, including classic concepts and algorithms such as spatial autocorrelation, heterogeneity, Kriging, and geographically weighted regression.

The system transforms this knowledge into a structured database and invokes it during AI evolution. Through intelligent retrieval and fusion (RAG - Fusion), it provides theoretically supported prompts for code generation.

In this way, AI can improve algorithms on the premise of "understanding geography," ensuring that the evolved models are both intelligent and reliable.

Figure 3. The framework of GeoKnowRAG

Case Study: Automatic Improvement of Kriging

Ordinary Kriging is the most classic spatial interpolation method in geostatistics and is widely used in environmental monitoring, climate analysis, and resource exploration.

However, its core algorithm was proposed a long time ago, and subsequent research has mostly involved "plug - in" combinations, such as combining with regression models to form regression kriging, with little new breakthrough in improving the structure of Kriging itself.

The brand - new Kriging model obtained by GeoEvolve

In the experiment, on the basis of retaining the core of Kriging, GeoEvolve automatically injected multiple improvements through evolution and knowledge guidance. Here are just a few examples:

Adaptive empirical variogram estimation

The original method uses fixed binning, which is easily affected by outliers.

GeoEvolve borrows statistical methods and introduces Silverman's binning rule, quantile binning, and trimmed mean to automatically determine the appropriate number of intervals.

Multi - starting point global fitting

Traditional fitting is prone to getting stuck in local optima.

GeoEvolve uses multi - starting point optimization + L1 or weighted least squares to ensure that the physical meaning of the parameters is reasonable (e.g., the range is non - negative).

Adaptive data transformation

When the data is highly skewed, GeoEvolve will automatically select the appropriate logarithmic transformation + offset to ensure a more reasonable residual distribution and more stable predictions.

Individually, these improvements may seem like just "incremental enhancements." However, when they are automatically combined and evolved by GeoEvolve and significantly improve the prediction accuracy in real - world experiments, they demonstrate powerful effects.

Compare the GeoEvolve - Kriging model with other Kriging models discovered by automated algorithms, and select the prediction of copper (Cu), lead (Pb), and zinc (Zn) in a mining area in Australia as the experimental scenario.

As shown in the table, the differences in the results are very obvious:

Original Kriging: Performs the worst, with the highest prediction error;

OpenEvolve - Kriging: Shows some improvement in some indicators but performs worse in zinc element prediction;

OpenEvolve with geographical knowledge prompts: Does not bring additional improvements, indicating that lack of targeted knowledge can hardly truly assist algorithm evolution;

GeoEvolve (version without knowledge base): Significantly outperforms OpenEvolve but is still inferior to the complete version;

Complete GeoEvolve - Kriging: Always performs the best, achieving the lowest RMSE and MAE in the prediction of all three metal elements.

Specifically, compared with OpenEvolve - Kriging, GeoEvolve reduces RMSE by 11.3% (Cu), 20.9% (Pb), and 13.5% (Zn); compared with the original Kriging, the reduction is even more significant, reaching 15.4%, 21.2%, and 13.0% respectively.

This clearly shows that the structured geographical knowledge base (GeoKnowRAG) plays a crucial role in algorithm evolution, enabling AI not only to write code but also to evolve a stronger interpolation model on the premise of "understanding geography."

Conclusion

The experimental results of GeoEvolve tell us that AI can do more than just modify code. It can also autonomously evolve stronger classic models under the guidance of geographical knowledge. This means that:

In the future, geographical modeling does not necessarily have to be "manually designed" by researchers and then modified based on LLMs;

We can fully automate the algorithm development process, allowing AI to continuously make trial - and - error attempts, absorb knowledge, and ultimately learn to think like a geographer;

GeoEvolve demonstrates a new path towards trustworthy GeoAI and opens up possibilities for the application of AI - for - Science in geographical science and sustainable development.

Perhaps in the near future, AI will not only be a tool but also a research collaborator.

References:

Luo, P., Lou, X., Zheng, Y., Zheng, Z. and Ermon, S., 2025. GeoEvolve: Automating Geospatial Model Discovery via Multi - Agent Large Language Models. arXiv preprint arXiv:2509.21593

This article is from the WeChat official account "New Intelligence Yuan", author: New Intelligence Yuan, editor: KingHZ, published by 36Kr with authorization.