Scientific Reasoning on GPQA (pass@1, pass@16)
48.1pass@1GRPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen3-8B-Base, Supervision Type=Labeled reward data2026.06 | 48.1 | 88.3 | |
| TTRL-CoCoVBackbone=Qwen3-8B-Base, Supervision Type=Label-free2026.06 | 43.6 | 89.9 | |
| Co-rewardingBackbone=Qwen3-8B-Base, Supervision Type=Label-free2026.06 | 41.5 | 75.6 | |
| GRPOBackbone=Qwen3-4B-Base, Supervision Type=Labeled reward data2026.06 | 40.5 | 83.4 | |
| TTRLBackbone=Qwen3-8B-Base, Supervision Type=Label-free2026.06 | 39.1 | 79.6 | |
| SCRLBackbone=Qwen3-8B-Base, Supervision Type=Label-free2026.06 | 38.2 | 82.5 | |
| TTRL-CoCoVBackbone=Qwen3-4B-Base, Supervision Type=Label-free2026.06 | 38.1 | 84 | |
| Co-rewardingBackbone=Qwen3-4B-Base, Supervision Type=Label-free2026.06 | 35.7 | 68.7 | |
| TTRLBackbone=Qwen3-4B-Base, Supervision Type=Label-free2026.06 | 35.6 | 74.9 | |
| BaseBackbone=Qwen3-8B-Base, Supervision Type=Label-free2026.06 | 35.3 | 88.9 | |
| SCRLBackbone=Qwen3-4B-Base, Supervision Type=Label-free2026.06 | 34.4 | 75.8 | |
| BaseBackbone=Qwen3-4B-Base, Supervision Type=Label-free2026.06 | 34.2 | 83.4 |