Software Engineering on SWE-Bench Verified (Reproduced/Reported)
23.6Reproduced Success RateProRL Agent-14B (RL)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ProRL Agent-14B (RL)Size=14B, Thinking mode=enabled2026.03 | 23.6 | — | |
| ProRL Agent-4B (RL)Size=4B2026.03 | 21.2 | — | |
| ProRL Agent-8B (RL)Size=8B, Thinking mode=enabled2026.03 | 18 | — | |
| Qwen3-14BSize=14B, Thinking mode=enabled2026.03 | 15.4 | — | |
| Qwen3-4B-Instruct-2507Size=4B2026.03 | 14.8 | — | |
| Qwen3-8BSize=8B, Thinking mode=enabled2026.03 | 9.6 | — | |
| SkyRL-Agent-14B-v0Size=14B2026.03 | — | 21.6 | |
| SkyRL-Agent-8B-v0Size=8B2026.03 | — | 9.4 |