Knowledge-Intensive Reasoning on Knowledge-Intensive Reasoning Suite (2Wiki, Bamb, HQA, MuSi, SimQA)
58.42Wiki ScoreQwen2.5-7B + SFT-then-RL
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Qwen2.5-7B + SFT-then-RLBackbone=Qwen2.5-7B, Training Protocol=SFT-then-RL2026.04 | 58.4 | 53.6 | 52.3 | 27 | 30.7 | |
| Llama3.1-8B + E3-TIR (Ours)Backbone=Llama3.1-8B, Training Protocol=E3-TIR2026.04 | 57.4 | 55.3 | 55.6 | 29.7 | 32.7 | |
| Qwen2.5-7B + E3-TIR (Ours)Backbone=Qwen2.5-7B, Training Protocol=E3-TIR2026.04 | 56.3 | 54.2 | 56 | 32.2 | 33.5 | |
| Qwen2.5-7B + Zero-RLBackbone=Qwen2.5-7B, Training Protocol=Zero-RL2026.04 | 55.7 | 49.6 | 50.1 | 24.7 | 31.5 | |
| ARPOBackbone=Qwen2.5-3B, Agent Category=Multi-Tool Integrated Agent2026.04 | 54.3 | 47.6 | 50 | 26.5 | 29.3 | |
| Llama3.1-8B + Zero-RLBackbone=Llama3.1-8B, Training Protocol=Zero-RL2026.04 | 54.3 | 50.9 | 49.6 | 28.5 | 31.6 | |
| Qwen2.5-3B + SFT-then-RLBackbone=Qwen2.5-3B, Training Protocol=SFT-then-RL2026.04 | 54.2 | 47.3 | 50.5 | 25.3 | 30.6 | |
| Llama3.1-8B + SFT-then-RLBackbone=Llama3.1-8B, Training Protocol=SFT-then-RL2026.04 | 53.8 | 55.4 | 53.8 | 30.2 | 33.1 | |
| E3-TIRBackbone=Qwen2.5-3B, Agent Category=Multi-Tool Integrated Agent2026.04 | 52.3 | 50.3 | 51.4 | 26.9 | 32.6 | |
| Qwen2.5-3B + E3-TIR (Ours)Backbone=Qwen2.5-3B, Training Protocol=E3-TIR2026.04 | 52.3 | 50.3 | 51.4 | 26.9 | 32.6 | |
| Tool-StarBackbone=Qwen2.5-3B, Agent Category=Multi-Tool Integrated Agent2026.04 | 48.6 | 44.1 | 52.4 | 27.8 | 30.3 | |
| Qwen2.5-3B + Zero-RLBackbone=Qwen2.5-3B, Training Protocol=Zero-RL2026.04 | 46.7 | 50.8 | 48.8 | 23.2 | 31.1 | |
| Search-R1Backbone=Qwen2.5-3B, Agent Category=LLM-based Search Agent2026.04 | 44.9 | 44.2 | 50.1 | 22.2 | 31.4 | |
| Tree-GRPOBackbone=Qwen2.5-3B, Agent Category=LLM-based Search Agent2026.04 | 44.1 | 43.7 | 50.5 | 26.7 | 29 | |
| ReCallBackbone=Qwen2.5-3B, Agent Category=Multi-Tool Integrated Agent2026.04 | 42.6 | 43.7 | 48.7 | 22.6 | 26.5 | |
| Qwen2.5-7B + Only SFTBackbone=Qwen2.5-7B, Training Protocol=Only SFT2026.04 | 33.3 | 37.9 | 37.1 | 14.6 | 23.2 | |
| Llama3.1-8B + Only SFTBackbone=Llama3.1-8B, Training Protocol=Only SFT2026.04 | 32.9 | 41.1 | 38.2 | 15.9 | 21.7 | |
| Qwen2.5-3B + Only SFTBackbone=Qwen2.5-3B, Training Protocol=Only SFT2026.04 | 30.3 | 35.2 | 35.8 | 16.3 | 19.5 | |
| Search-o1Backbone=Qwen2.5-3B, Agent Category=LLM-based Search Agent2026.04 | 27.4 | 38.5 | 29.5 | 12.5 | 10.8 | |
| ToRLBackbone=Qwen2.5-3B, Agent Category=LLM-based Math Agent2026.04 | 24 | 25.4 | 17.9 | 8.3 | 6.5 | |
| Qwen2.5-3B-InstructBackbone=Qwen2.5-3B2026.04 | 19.2 | 27.5 | 11.5 | 6.3 | 5.9 | |
| Qwen2.5-3B-InstructBackbone=Qwen2.5-3B, Training Protocol=Instruct2026.04 | 19.2 | 27.5 | 11.5 | 6.3 | 5.9 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B, Training Protocol=Instruct2026.04 | 18.2 | 29.6 | 15.3 | 7.8 | 6.4 | |
| Llama3.1-8B-InstructBackbone=Llama3.1-8B, Training Protocol=Instruct2026.04 | 16.9 | 38.4 | 12.1 | 13.3 | 6.3 | |
| SimpleTIRBackbone=Qwen2.5-3B, Agent Category=LLM-based Math Agent2026.04 | 16.1 | 24.2 | 16.5 | 7.2 | 8.8 |