Reasoning on GPQA Diamond (pass@4 accuracy)
67.17Pass@4 AccuracySAGC
Evaluation Results
| Method | Links | |
|---|---|---|
| SAGCBackbone=Llama3.2-3b, Algorithm=DAPO, Group Size=Dynamic2026.06 | 67.17 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=GRPO, Group Size=162026.06 | 63.64 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=DAPO, Group Size=42026.06 | 62.63 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=DAPO, Group Size=82026.06 | 62.63 | |
| SAGCBackbone=Llama3.2-3b, Algorithm=GRPO, Group Size=Dynamic2026.06 | 62.12 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=GRPO, Group Size=42026.06 | 61.11 | |
| SAGCBackbone=Qwen2.5-3b, Algorithm=GRPO, Group Size=Dynamic2026.06 | 61.11 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=DAPO, Group Size=162026.06 | 61.11 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=GRPO, Group Size=162026.06 | 60.61 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=GRPO, Group Size=82026.06 | 57.58 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=GRPO, Group Size=82026.06 | 57.07 | |
| OpenRLHFBackbone=Llama3.2-3b, Algorithm=GRPO, Group Size=42026.06 | 57.07 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=DAPO, Group Size=82026.06 | 56.06 | |
| SAGCBackbone=Qwen2.5-3b, Algorithm=DAPO, Group Size=Dynamic2026.06 | 56.06 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=DAPO, Group Size=42026.06 | 55.05 | |
| OpenRLHFBackbone=Qwen2.5-3b, Algorithm=DAPO, Group Size=162026.06 | 55.05 |