Knowledge-intensive reasoning on Bamboogle (F1, EM, TC)
60.4F1 ScoreEAPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EAPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 60.4 | 48.8 | 2.17 | |
| GRPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 55.6 | 44.8 | 2.8 | |
| AEPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 55.6 | 46.4 | 4.16 | |
| Reinforce++Backbone Model=Qwen2.5-7B-Instruct2026.06 | 54.2 | 42.4 | 1.94 | |
| ARPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 51.8 | 40 | 1.93 | |
| ARPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 48.1 | 38.4 | 2.13 | |
| EAPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 48 | 40 | 1.9 | |
| ToolStarBackbone Model=Qwen2.5-7B-Instruct2026.06 | 48 | 36 | 4.28 | |
| GRPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 44.8 | 34.5 | 2.57 | |
| ToolStarBackbone Model=Qwen2.5-3B-Instruct2026.06 | 40.7 | 30.4 | 2.31 | |
| Reinforce++Backbone Model=Qwen2.5-3B-Instruct2026.06 | 39.4 | 30.6 | 2.27 | |
| BaseBackbone Model=Qwen2.5-7B-Instruct2026.06 | 20.1 | 12 | — | |
| TIRBackbone Model=Qwen2.5-3B-Instruct2026.06 | 16.4 | 10 | 1.32 | |
| BaseBackbone Model=Qwen2.5-3B-Instruct2026.06 | 11.7 | 7.4 | — | |
| TIRBackbone Model=Qwen2.5-7B-Instruct2026.06 | 11.2 | 7.3 | 2.3 |