This paper seems pretty relevant. Just skimmed through, and it looks like they tackle the same problem setting with RL but with a special

jump

tool. https://arxiv.org/abs/2512.20957

Interesting, their results seem a bit poor though?

I think they use swe-smith for training and swe-bench-verified + swe-bench-pro for validation. Table 1 in the screenshot above seems to be swe-bench-verified I think. They also have a second table for results on a subset of swe-bench-pro.

I started an issue where we can add related work here! https://github.com/OpenHands/agentic-code-search-oss/issues/54