Downstream Policy Performance on Arena-Hard v2.0
33.9Win RateGRPO + Improved RM
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO + Improved RMPolicy Model=Qwen3-4B-Instruct-25072026.05 | 33.9 | |
| GRPO + SAVEPolicy Model=Qwen3-4B-Instruct-25072026.05 | 33.5 | |
| GRPO + R2MPolicy Model=Qwen3-4B-Instruct-25072026.05 | 32.6 | |
| SFTPolicy Model=Qwen3-4B-Instruct-25072026.05 | 31.9 | |
| GRPOPolicy Model=Qwen3-4B-Instruct-25072026.05 | 30.2 | |
| REINFORCE++Policy Model=Qwen3-4B-Instruct-25072026.05 | 28.6 | |
| PRIMEPolicy Model=Qwen3-4B-Instruct-25072026.05 | 22.7 | |
| GRPO + Improved RMPolicy Model=Qwen2.5-3B-Instruct2026.05 | 2.6 | |
| GRPOPolicy Model=Qwen2.5-3B-Instruct2026.05 | 2.2 | |
| GRPO + SAVEPolicy Model=Qwen2.5-3B-Instruct2026.05 | 2.2 | |
| REINFORCE++Policy Model=Qwen2.5-3B-Instruct2026.05 | 2.1 | |
| SFTPolicy Model=Qwen2.5-3B-Instruct2026.05 | 2 | |
| GRPO + R2MPolicy Model=Qwen2.5-3B-Instruct2026.05 | 1.9 | |
| PRIMEPolicy Model=Qwen2.5-3B-Instruct2026.05 | 1.3 |