Tree search and/or agentic behavior

[amueller]

There's this very interesting recent work on program generation via tree search from google research: https://arxiv.org/pdf/2509.06503

In a sense, it's very similar to alpha-evolve, but there are two key differences that I can see:

• The search is done using tree search instead of the island based / MAP-elites base search
• Generation is done with an agent, instead of one-shot.

Would you be interested in incorporating either of both of these features in this repo?

[codelion]

I had a quick look at the paper, that approach may be a better addition to OptiLLM - https://github.com/codelion/optillm v/s in this repo. We had implemented mcts in OptiLLM last year.

It would be interesting to see how OpenEvolve can solve some of the benchmarks used in the paper with the island based search as implemented right now. There has been discussion in the past on adding more types of evolutionary approaches in (see #186) but that will require careful redesign of OpenEvolve to support a new interface that could implement different approaches.

[BradKML]

We need to cover this as well to make the new version of OpenEvolve to be on par https://github.com/SakanaAI/ShinkaEvolve

[amueller]

@BradKML ShinkaEvolve doesn't seem to support tree search, right?

[BradKML]

@amueller I think they do tree-like searching methods? They are definitely agentic tho (so they build their own tools)