Multi-Task Reasoning on GPQA Diamond
55Pass@1Standard RLVR
Evaluation Results
| Method | Links | |
|---|---|---|
| Standard RLVRBackbone=Qwen3-14B-Base2026.02 | 55 | |
| Composition-RLBackbone=Qwen3-30B-A3B-Base2026.02 | 54.6 | |
| Composition-RLBackbone=Qwen3-14B-Base2026.02 | 54.2 | |
| Standard RLVRBackbone=Qwen3-30B-A3B-Base2026.02 | 50.7 | |
| Composition-RLBackbone=Qwen3-8B-Base2026.02 | 48.9 | |
| Composition-RLBackbone=Qwen3-4B-Base, Curriculum=Depth 1 + 2 + 32026.02 | 48.5 | |
| Composition-RLBackbone=Qwen3-4B-Base, Curriculum=Depth 1 + 22026.02 | 48.3 | |
| Standard RLVRBackbone=Qwen3-8B-Base2026.02 | 48.2 | |
| Composition-RLBackbone=Qwen3-4B-Base2026.02 | 46.3 | |
| RULEREASONER-8B + DADSBackbone=Qwen3-8B, Algorithm=DADS2025.06 | 44.9 | |
| Standard RLVRBackbone=Qwen3-4B-Base2026.02 | 43.7 | |
| DCRLBackbone=Qwen3-8B-Base2026.03 | 37.9 | |
| DCRLBackbone=Qwen3-4B-Base2026.03 | 34.3 | |
| DCRLBackbone=Llama3.2-3B-Instruct2026.03 | 24.7 | |
| Qwen3-8B-BaseBackbone=Qwen3-8B2025.06 | 16.6 |