Multi-Task on GPQA
41.7Avg@32 AccuracyIBTPO
Evaluation Results
| Method | Links | |
|---|---|---|
| IBTPOBase Model=Qwen3-8B-Base2026.05 | 41.7 | |
| TreeRLBase Model=Qwen3-8B-Base2026.05 | 39.8 | |
| TreePOBase Model=Qwen3-8B-Base2026.05 | 39.2 | |
| GRPO w/ Entropy Reg†Base Model=Qwen3-8B-Base2026.05 | 39.1 | |
| IBROBase Model=Qwen3-8B-Base2026.05 | 38.5 | |
| Vanilla GRPOBase Model=Qwen3-8B-Base2026.05 | 38.1 | |
| GRPO w/ Clip-higherBase Model=Qwen3-8B-Base2026.05 | 37.8 | |
| IBTPOBase Model=Qwen3-1.7B-Base2026.05 | 29 | |
| TreeRLBase Model=Qwen3-1.7B-Base2026.05 | 26.8 | |
| Vanilla GRPOBase Model=Qwen3-1.7B-Base2026.05 | 26.5 | |
| TreePOBase Model=Qwen3-1.7B-Base2026.05 | 26.2 | |
| Initial ModelBase Model=Qwen3-8B-Base2026.05 | 26.1 | |
| GRPO w/ Clip-higherBase Model=Qwen3-1.7B-Base2026.05 | 24.8 | |
| GRPO w/ Entropy Reg†Base Model=Qwen3-1.7B-Base2026.05 | 24.7 | |
| IBROBase Model=Qwen3-1.7B-Base2026.05 | 24.7 | |
| Initial ModelBase Model=Qwen3-1.7B-Base2026.05 | 21.2 |