When they tested on bugs not in SWE-Bench, the success rate dropped to 57‑71% on random items, and 50‑68% on fresh issues created after the benchmark snapshot. I’m surprised they did that well.
After the benchmark snapshot. Could still be before LLM training data cut off, or available via RAG.
edit: For a fair test you have to use git issues that had not been resolved yet by a human.
This is how these fuckers talk, all of the time. Also see Sam Altman's not-quite-denials of training on Scarlett Johansson's voice: they just asserted that they had hired a voice actor, but didn't deny training on actual Scarlett Johansson's voice. edit: because anyone with half a brain knows that not only did they train on her actual voice, they probably gave it and their other pirated movie soundtracks massively higher weighting, just as they did for books and NYT articles.
Anyhow, I fully expect that by now they just use everything they can to cheat benchmarks, up to and including RAG from solutions past the training dataset cut off date. With two of the paper authors being from Microsoft itself, expect that their "fresh issues" are gamed too.