Task Performance on GDPVal 44 tasks (held-out)
0.81Mean ReturnOracle Rubric (Baseline)
Evaluation Results
| Method | Links | |
|---|---|---|
| Oracle Rubric (Baseline)N (Number of rollouts)=82025.12 | 0.81 | |
| GSPO ModelN (Number of rollouts)=82025.12 | 0.74 | |
| Oracle Rubric (Baseline)N (Number of rollouts)=12025.12 | 0.7 | |
| SFT ModelN (Number of rollouts)=82025.12 | 0.68 | |
| GSPO ModelN (Number of rollouts)=12025.12 | 0.62 | |
| SFT ModelN (Number of rollouts)=12025.12 | 0.59 | |
| No Rubric (LLM-Judged)N (Number of rollouts)=12025.12 | 0.58 | |
| No Rubric (LLM-Judged)N (Number of rollouts)=82025.12 | 0.58 |