I heard about the ShinkaEvolve on a podcast where the guest had used it to evolve an agent harness for a less capable model.
I ended up borrowing the ideas from it for one of my own personal projects.
I ended up borrowing the ideas from it for one of my own personal projects.
SGD generates a stronger learning signal,is more efficient, and scales better. Using it end-to-end makes it stronger yet.
Yet somehow mixing in a weaker blunt evolution stage improves the result?