not really! it is a different setup (i.e., the setup we have in
SWE-Gym -- basically similar to the right part of figure i shared), where we run multiple independent runs on the same problem, and trained a verifier/reward model to help pick the best trajectory as the
final solution. I'm still wrapping things up to release the reward model we used for this 🙂
This is the pre-cursor for this planned feature that i'm gonna work on next: for a problem, we first ask an agent to solve it, ideally, we can have the reward model telling us IF the problem has been solved or not. If it gives a low probability, we can ask the agent to do it again in a fresh new session (maybe with some lessons learned/summarized from previous trajectory).
https://github.com/All-Hands-AI/OpenHands/issues/2221