Technical problem-solving on Arena Hard
52.3Win RateGRPO (RLVRR)
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO (RLVRR)#Data=10K2026.01 | 52.3 | |
| SFT#Data=100K2026.01 | 51 | |
| GRPO (RM)#Data=10K2026.01 | 50.7 | |
| GRPO (RLPR)#Data=10K2026.01 | 48.6 | |
| GRPO (GRM)#Data=10K2026.01 | 48.5 | |
| DPO#Data=10K2026.01 | 48.1 | |
| SFT#Data=10K2026.01 | 46.9 | |
| GRPO (BLEU)#Data=10K2026.01 | 39.6 | |
| Instruct#Data=n/a2026.01 | 34.3 | |
| GRPO (Random)#Data=10K2026.01 | 6.7 |