Question Answering on NQ (F1 Score)
78.8F1 Score (NQ)Tri-RAG
Evaluation Results
| Method | Links | |
|---|---|---|
| Tri-RAGGenerator Backbone=Qwen-Max2026.04 | 78.8 | |
| Tri-RAGGenerator Backbone=Qwen-Long2026.04 | 78.7 | |
| AflowGenerator Backbone=Qwen-Long2026.04 | 74.6 | |
| AflowGenerator Backbone=Qwen-Max2026.04 | 73.9 | |
| SafeChain2026.04 | 73.9 | |
| Tri-RAGGenerator Backbone=Gemma-2-27B2026.04 | 72.5 | |
| STAR-12026.04 | 72.3 | |
| R1-ALT2026.04 | 72 | |
| No train2026.04 | 71.7 | |
| Tri-RAGGenerator Backbone=GPT-4o-mini2026.04 | 69.8 | |
| Tri-RAGGenerator Backbone=GPT-4.1-nano2026.04 | 69.8 | |
| AflowGenerator Backbone=Gemma-2-27B2026.04 | 69.4 | |
| Self-ActGenerator Backbone=Qwen-Long2026.04 | 68.9 | |
| AflowGenerator Backbone=GPT-4o-mini2026.04 | 67.9 | |
| Self-ActGenerator Backbone=Qwen-Max2026.04 | 67.8 | |
| AflowGenerator Backbone=GPT-4.1-nano2026.04 | 66.5 | |
| Self-ActGenerator Backbone=Gemma-2-27B2026.04 | 65.3 | |
| Self-ActGenerator Backbone=GPT-4o-mini2026.04 | 62.4 | |
| Self-ActGenerator Backbone=GPT-4.1-nano2026.04 | 59.8 | |
| ReActGenerator Backbone=Qwen-Max2026.04 | 59.8 | |
| ReActGenerator Backbone=Gemma-2-27B2026.04 | 57.8 | |
| ReActGenerator Backbone=Qwen-Long2026.04 | 56.3 | |
| PiCABackbone=Qwen2.5-7B Instruct2026.05 | 55.3 | |
| DirectGenerator Backbone=Qwen-Max2026.04 | 53.3 | |
| TIPSBackbone=Qwen2.5-7B Instruct2026.05 | 53.2 | |
| GraphRAG-RouterSystem=Our Method2026.03 | 52.5 | |
| PiCABackbone=Qwen2.5-3B Instruct2026.05 | 52.1 | |
| TIPSBackbone=Qwen2.5-3B Instruct2026.05 | 51.8 | |
| DirectGenerator Backbone=Gemma-2-27B2026.04 | 51.3 | |
| MT-PPOBackbone=Qwen2.5-7B Instruct2026.05 | 50.6 | |
| Search-R1Backbone=Qwen2.5-7B Instruct2026.05 | 49.6 | |
| Router-R1System=RL-based Agent2026.03 | 49.5 | |
| ReActGenerator Backbone=GPT-4o-mini2026.04 | 48.9 | |
| GraphSearchSystem=Training-free Agent2026.03 | 48.3 | |
| MT-PPOBackbone=Qwen2.5-3B Instruct2026.05 | 47.8 | |
| Search-o1System=Training-free Agent2026.03 | 47.6 | |
| Search-R1System=RL-based Agent2026.03 | 47.3 | |
| HippoRAG2System=RAG-Based2026.03 | 46.5 | |
| IGPOBackbone=Qwen2.5-7B-Instruct2025.10 | 46.4 | |
| Search-R1Backbone=Qwen2.5-3B Instruct2026.05 | 45.7 | |
| ReActGenerator Backbone=GPT-4.1-nano2026.04 | 45.4 | |
| LinearRAGSystem=RAG-Based2026.03 | 43.6 | |
| DirectGenerator Backbone=Qwen-Long2026.04 | 43.4 | |
| GPT-4o (2024-11-20)Version=2024-11-202025.08 | 42.74 | |
| GraphRAGSystem=RAG-Based2026.03 | 42.3 | |
| GSPOBackbone=Qwen2.5-7B-Instruct2025.10 | 41.5 | |
| Graph-R1System=RL-based Agent2026.03 | 41.5 | |
| RLOOBackbone=Qwen2.5-7B-Instruct2025.10 | 40.7 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2025.10 | 40.3 | |
| Vanilla RAGSystem=RAG-Based2026.03 | 40.1 | |
| CoTSystem=Basic LLM2026.03 | 39.9 | |
| PPOBackbone=Qwen2.5-7B-Instruct2025.10 | 38.7 | |
| RAPTORSystem=RAG-Based2026.03 | 38.2 | |
| HyperGraphRAGSystem=RAG-Based2026.03 | 37.5 | |
| DirectGenerator Backbone=GPT-4o-mini2026.04 | 36.4 | |
| Reinforce++Backbone=Qwen2.5-7B-Instruct2025.10 | 34.3 | |
| DirectGenerator Backbone=GPT-4.1-nano2026.04 | 33.7 | |
| Llama + LoRRBackbone=Llama3.2-3B, Optimization=LoRR2025.08 | 18.24 | |
| Llama + DPOBackbone=Llama3.2-3B, Optimization=DPO2025.08 | 18.03 | |
| Qwen + LoRRBackbone=Qwen2.5-3B, Optimization=LoRR2025.08 | 17.2 | |
| Qwen (Base)Backbone=Qwen2.5-3B, Mode=Base2025.08 | 17.18 | |
| Qwen + DPOBackbone=Qwen2.5-3B, Optimization=DPO2025.08 | 16.83 | |
| Direct InferSystem=Basic LLM2026.03 | 16.5 | |
| Llama (Base)Backbone=Llama3.2-3B, Mode=Base2025.08 | 14.75 |