Mathematics on OlymMATH-Hard
3.2Pass@1GSPO RL with novelty bonus
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GSPO RL with novelty bonustemperature=0.6, rollouts=32, Backbone=OLMo3-7B-Instruct-SFT2026.05 | 3.2 | 36 | |
| GSPO RL (vanilla)temperature=0.6, rollouts=32, Backbone=OLMo3-7B-Instruct-SFT2026.05 | 2.8 | 29 | |
| Puzzle SFT (rejection-sampled DSR traces)temperature=0.6, rollouts=32, Backbone=OLMo3-7B-Instruct-SFT2026.05 | 2.7 | 23 | |
| Base (OLMo3-7B-Instruct-SFT)temperature=0.6, rollouts=322026.05 | 1.5 | 16 |