An AI agent called SWE-Doctor is fixing software bugs at a rate that should make traditional patch-generation tools nervous.
Researchers built SWE-Doctor to address a specific failure mode in LLM-based software engineering agents: bug reproduction tests, while useful for validating patches after the fact, were actively hurting patch generation when used naively. The team found two culprits. Tests that fail when they should pass often capture only one symptom of a bug, leading agents to ship partial fixes. Tests that fail when they should fail are unreliable targets altogether. SWE-Doctor's answer is to generate multiple tests covering different behavioral requirements stated in the issue report, execute and debug them to build what the paper calls "runtime-grounded diagnosis records," then feed those records into patch generation alongside localization information gathered during testing.
The results are hard to dismiss: 75.7% resolution on SWE-bench Verified and 59.4% on the harder SWE-bench Pro, across five different LLM backends. On SWE-bench Pro, that represents an 8.0 to 8.9 percentage-point gain over the baseline agents - meaningful headroom in a space where single-digit improvements are usually celebrated as breakthroughs.
SWE-bench Pro is the newer, deliberately trickier sibling of Verified, designed specifically because labs were starting to saturate the original benchmark - so clearing 59.4% there carries more signal than a high score on the older test would.