Science Question Answering on GPQA-Diamond Science N=198
41.4AccuracyRVPO-explicit (β=1.0/0.5)
Evaluation Results
| Method | Links | |
|---|---|---|
| RVPO-explicit (β=1.0/0.5)Size=14B, Checkpoint=Best Ckpt, β=1.0/0.52026.05 | 41.4 | |
| GRPO (Explicit)Size=14B, Checkpoint=Best Ckpt2026.05 | 40.4 | |
| GRPO (Implicit)Size=14B, Checkpoint=Best Ckpt2026.05 | 39.4 | |
| GDPOSize=14B, Checkpoint=Best Ckpt2026.05 | 39.4 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=14B, Checkpoint=Best Ckpt, k=0.5→2.0 / 1.0→2.02026.05 | 38.4 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=14B, Checkpoint=Final (300), k=0.5→2.0 / 1.0→2.02026.05 | 38.4 | |
| GDPOSize=14B, Checkpoint=Final (300)2026.05 | 37.4 | |
| RVPO-min (k=∞)Size=14B, Checkpoint=Best Ckpt, k=∞2026.05 | 36.9 | |
| GRPO (Implicit)Size=14B, Checkpoint=Final (300)2026.05 | 35.9 | |
| RVPO-explicit (β=1.0/0.5)Size=14B, Checkpoint=Final (300), β=1.0/0.52026.05 | 35.9 | |
| RVPO-min (k=∞)Size=3B, Checkpoint=Best Ckpt, k=∞2026.05 | 34.8 | |
| GRPO (Explicit)Size=3B, Checkpoint=Best Ckpt2026.05 | 34.3 | |
| GRPO (Explicit)Size=3B, Checkpoint=Final (300)2026.05 | 34.3 | |
| GRPO (Explicit)Size=7B, Checkpoint=Best Ckpt2026.05 | 34.3 | |
| GRPO (Explicit)Size=7B, Checkpoint=Final (300)2026.05 | 34.3 | |
| GRPO (Implicit)Size=3B, Checkpoint=Best Ckpt2026.05 | 33.8 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=7B, Checkpoint=Best Ckpt, k=0.5→2.0 / 1.0→2.02026.05 | 33.8 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=7B, Checkpoint=Final (300), k=0.5→2.0 / 1.0→2.02026.05 | 33.8 | |
| GRPO (Explicit)Size=14B, Checkpoint=Final (300)2026.05 | 33.8 | |
| GDPOSize=7B, Checkpoint=Best Ckpt2026.05 | 32.3 | |
| RVPO-min (k=∞)Size=7B, Checkpoint=Best Ckpt, k=∞2026.05 | 32.3 | |
| GRPO (Implicit)Size=7B, Checkpoint=Best Ckpt2026.05 | 31.8 | |
| GDPOSize=7B, Checkpoint=Final (300)2026.05 | 31.8 | |
| RVPO-explicit (β=1.0/0.5)Size=7B, Checkpoint=Best Ckpt, β=1.0/0.52026.05 | 31.8 | |
| RVPO-explicit (β=1.0/0.5)Size=7B, Checkpoint=Final (300), β=1.0/0.52026.05 | 31.8 | |
| GRPO (Implicit)Size=3B, Checkpoint=Final (300)2026.05 | 31.3 | |
| RVPO-explicit (β=1.0/0.5)Size=3B, Checkpoint=Best Ckpt, β=1.0/0.52026.05 | 31.3 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=3B, Checkpoint=Best Ckpt, k=0.5→2.0 / 1.0→2.02026.05 | 31.3 | |
| GDPOSize=3B, Checkpoint=Best Ckpt2026.05 | 30.8 | |
| RVPO-explicit (β=1.0/0.5)Size=3B, Checkpoint=Final (300), β=1.0/0.52026.05 | 30.8 | |
| RVPO-min (k=∞)Size=7B, Checkpoint=Final (300), k=∞2026.05 | 30.8 | |
| RVPO (k=0.5→2.0 / 1.0→2.0)Size=3B, Checkpoint=Final (300), k=0.5→2.0 / 1.0→2.02026.05 | 30.3 | |
| GDPOSize=3B, Checkpoint=Final (300)2026.05 | 29.3 | |
| RVPO-min (k=∞)Size=14B, Checkpoint=Final (300), k=∞2026.05 | 29.3 | |
| RVPO-min (k=∞)Size=3B, Checkpoint=Final (300), k=∞2026.05 | 28.3 | |
| GRPO (Implicit)Size=7B, Checkpoint=Final (300)2026.05 | 28.3 |