FeedDiscussion
thats interesting, we're also using gepa alongside other algoriths. I know of anbuilt an open source repo with these that helps u analyze your eval scores and data to further optimize your prompt automatically -https://github.com/future-agi/future-agi
iterating on prompts is such a bottleneck, really cool to see this approach.
unblgamesg
Really interesting approach. I like the idea of keeping a Pareto front instead of optimizing toward a single prompt, since that seems much less likely to overfit to the latest failure case. town of salem