Coding Agent on SWE-bench Verified (single-run, single-patch, test)
51.6Resolution RateQwen3-Coder
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3-CoderScale=30B2026.05 | 51.6 | — | — | — | |
| Sync. OPDScale=8B2026.05 | 41.8 | 1 | 1.6 | 0 | |
| Vanilla GRPOScale=8B2026.05 | 40.2 | 0.92 | — | — | |
| SA-SWEScale=32B2026.05 | 39.4 | — | — | — | |
| f-OPDScale=8B2026.05 | 39.4 | 1.46 | 2.6 | 0 | |
| Vanilla SFTScale=8B2026.05 | 39.1 | 2.18 | — | 0 | |
| DeepSWEScale=32B2026.05 | 36.4 | — | — | — | |
| Async. OPD + hard refreshScale=8B2026.05 | 35.1 | 1.15 | 6.6 | 1 | |
| Async. OPD + lag-only weightingScale=8B2026.05 | 34.9 | 1.54 | 7.6 | 2 | |
| Async. OPDScale=8B2026.05 | 26.8 | 1.61 | 12.1 | 3 | |
| SWE-GymScale=32B2026.05 | 20.6 | — | — | — |