Feature Proposal: New Symbolic Regression Example for OpenEvolve

[linhaowei1]

Hi @codelion,

Firstly, fantastic work on OpenEvolve! As someone working in LLM-assisted scientific discovery, I've been very impressed with its capabilities.

I recently applied OpenEvolve to a new symbolic regression benchmark (link to arXiv paper: 2504.10415), which is an ICML 2025 spotlight paper, and found it performs very effectively. My implementation is available in my fork here.

I believe an open-sourced, unified interface for running SR benchmarks like this within OpenEvolve could greatly benefit the scientific discovery community. Would you be open to merging well-prepared examples into the main repository?

If this is of interest, I'd be happy to refine my current implementation (e.g., remove personal file paths, add a comprehensive README) and submit a pull request. I understand if this isn't a current priority, given that symbolic regression isn't the primary showcase of the AlphaEvolve paper.

Should this proposal be well-received, I would be keen to contribute further scientific discovery examples.

Looking forward to your thoughts.

Best regards,
Haowei Lin

[codelion]

Hi @linhaowei1 that sounds good. We can try to keep the structure of the existing examples with evaluate.py and config.yaml. As long as we can run the examples consistently it should be a great addition to the repo. Let me know if you need any help.
Thanks.

[linhaowei1]

Thanks, @codelion! Glad to hear this sounds good.

Regarding the structure, my current setup for the benchmark places evaluate.py and config.yaml within separate subdirectories for each task (e.g., examples/symbolic_regression/problems/chem_react/CRK0). A script in a higher-level directory then runs OpenEvolve for these individual tasks.

Does this approach align with what you have in mind, or would you prefer a different organization? I'm happy to adjust.

Perhaps establishing a simple template for benchmark organization could be helpful for standardizing this and future contributions. What are your thoughts on that?

[codelion]

I think this structure is fine, as long as we can run the full benchmark from the /symbolic_regresions folder we should be fine. W can use the README int hat directory to document the example and results etc.

[linhaowei1]

Got it. I'll prepare the pull request with these points in mind and will have it ready soon.

[linhaowei1]

Hi @codelion ,

The PR is #17 . Feel free to take a look.