Code Generation on Code (test)
36.5AccuracyNon-hacking
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Non-hackingBase Model=Qwen2.5-3B-Instruct2026.04 | 36.5 | — | |
| RFT+GRIFTBase Model=Qwen2.5-3B-Instruct, Checkpoint Step=202026.04 | 23.3 | 32.8 | |
| Starting-pointBase Model=Qwen2.5-3B-Instruct, Checkpoint Step=202026.04 | 19.6 | 49.5 | |
| No-interventionBase Model=Qwen2.5-3B-Instruct2026.04 | 16.2 | 81.9 | |
| RFT+RandomBase Model=Qwen2.5-3B-Instruct, Checkpoint Step=202026.04 | 15.9 | 54.6 | |
| RFT+TraceBase Model=Qwen2.5-3B-Instruct, Checkpoint Step=202026.04 | 15.4 | 67.2 |