Knowledge-intensive reasoning on 2WikiMultihopQA (F1, EM, TC)
58.6F1 ScoreEAPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EAPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 58.6 | 51.5 | 1.71 | |
| EAPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 53.4 | 45.5 | 2.04 | |
| AEPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 48.8 | 42.5 | 4.95 | |
| ARPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 45.4 | 40.5 | 2.37 | |
| GRPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 45 | 37.1 | 2.95 | |
| ARPOBackbone Model=Qwen2.5-7B-Instruct2026.06 | 44.9 | 40 | 2.02 | |
| Reinforce++Backbone Model=Qwen2.5-7B-Instruct2026.06 | 44.2 | 36.5 | 2.09 | |
| GRPOBackbone Model=Qwen2.5-3B-Instruct2026.06 | 42.1 | 38.9 | 2.41 | |
| ToolStarBackbone Model=Qwen2.5-7B-Instruct2026.06 | 42.1 | 30.9 | 4.89 | |
| Reinforce++Backbone Model=Qwen2.5-3B-Instruct2026.06 | 39.7 | 37.4 | 2.62 | |
| ToolStarBackbone Model=Qwen2.5-3B-Instruct2026.06 | 31.5 | 27.5 | 2.72 | |
| TIRBackbone Model=Qwen2.5-3B-Instruct2026.06 | 14.1 | 9.4 | 1.6 | |
| TIRBackbone Model=Qwen2.5-7B-Instruct2026.06 | 12.1 | 5.1 | 2.91 | |
| BaseBackbone Model=Qwen2.5-7B-Instruct2026.06 | 11.6 | 5 | — | |
| BaseBackbone Model=Qwen2.5-3B-Instruct2026.06 | 9.4 | 6.5 | — |