Reasoning on GPQA Diamond (Pass@1, FS, N/T, D/T, Repro)
66.8Pass@1 AccuracyRED
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| REDModel=Qwen3-32B-Thinking2026.04 | 66.8 | 2.3 | 3.2 | 2.3 | 2 | |
| S-GRROModel=Qwen3-32B-Thinking2026.04 | 66.5 | 4.6 | 5.3 | 3.8 | 5.6 | |
| REDModel=R1-Llama-70B2026.04 | 66.2 | 2.9 | 3.5 | 2.8 | 3 | |
| RL + LPModel=Qwen3-32B-Thinking2026.04 | 66 | 4.4 | 5.2 | 3.5 | 5.3 | |
| VanillaModel=Qwen3-32B-Thinking2026.04 | 65.3 | 5.1 | 5.6 | 4 | 6.1 | |
| S-GRROModel=R1-Llama-70B2026.04 | 65.2 | 5.9 | 5.7 | 4.6 | 8.6 | |
| RL + LPModel=R1-Llama-70B2026.04 | 65 | 5.7 | 5.6 | 4.4 | 8.3 | |
| DEERModel=Qwen3-32B-Thinking2026.04 | 64.8 | 4.7 | 5.4 | 4.1 | 6.2 | |
| VanillaModel=R1-Llama-70B2026.04 | 64.5 | 6.5 | 6.2 | 4.8 | 9.4 | |
| DEERModel=R1-Llama-70B2026.04 | 63.1 | 5.9 | 5.9 | 4.9 | 9.6 | |
| REDModel=R1-Qwen-32B2026.04 | 61.8 | 2.4 | 3.5 | 2.7 | 3 | |
| S-GRROModel=R1-Qwen-32B2026.04 | 61.3 | 4.7 | 5.6 | 4.2 | 6.6 | |
| VanillaModel=R1-Qwen-32B2026.04 | 60.8 | 5.3 | 5.9 | 4.4 | 7.1 | |
| RL + LPModel=R1-Qwen-32B2026.04 | 60.4 | 4.6 | 5.5 | 3.9 | 6.4 | |
| DEERModel=R1-Qwen-32B2026.04 | 58.9 | 4.9 | 5.7 | 4.4 | 7.3 |