Can LLMs fix each other's bugs?

How well can a language model repair a bug written by a different model — or by a human? This automated-program-repair benchmark puts models on both sides of the problem. Each model injects faults into twenty curated SWE-bench Lite instances, singly or compounded on top of the existing human-written bug, and then every model — including the injector — attempts the repair, producing an injector-by-repairer matrix that measures self-bias alongside standard pass@k.

The bugs are generated at run time and have to earn their place: an injection only counts if its patch applies and the official SWE-bench Docker harness confirms the tests now fail, and each repair is re-diffed against the original repository so the unmodified harness can judge it. Around that sits the unglamorous machinery of a multi-provider LLM experiment — defensive patch extraction from free-form model output, per-model cost accounting with a no-API-call estimate mode, and resumable phases that skip work already paid for. An earlier arm of the same design runs on QuixBugs, with bootstrap confidence intervals and significance tests built in for publication-grade analysis.