Question Answering on NQ (test) (Unreliability Rate)
59.1Unreliability RateKbSD
Evaluation Results
| Method | Links | |
|---|---|---|
| KbSDBackbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 59.1 | |
| KbPOBackbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 62.3 | |
| KbSDBackbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 64.4 | |
| Search-R1Backbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 65.27 | |
| ReSearchBackbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 65.39 | |
| ARPOBackbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 66.71 | |
| KbPOBackbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 68.5 | |
| ReACTBackbone=Qwen2.5-7B, Paradigm=Agentic RAG2026.06 | 69.99 | |
| Search-R1Backbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 70.41 | |
| TCRAGBackbone=Qwen2.5-7B, Paradigm=Agentic RAG2026.06 | 70.99 | |
| IRCOTBackbone=Qwen2.5-3B, Paradigm=Agentic RAG2026.06 | 72.14 | |
| ReACTBackbone=Qwen2.5-3B, Paradigm=Agentic RAG2026.06 | 72.81 | |
| AEPOBackbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 73.24 | |
| Mem1Backbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 73.62 | |
| ReSearchBackbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 76.31 | |
| COTBackbone=Qwen2.5-7B, Paradigm=No RAG2026.06 | 77.53 | |
| FL-RAGBackbone=Qwen2.5-3B, Paradigm=Naive RAG2026.06 | 78.07 | |
| TCRAGBackbone=Qwen2.5-3B, Paradigm=Agentic RAG2026.06 | 79.26 | |
| FS-RAGBackbone=Qwen2.5-3B, Paradigm=Naive RAG2026.06 | 80.16 | |
| BaseBackbone=Qwen2.5-7B, Paradigm=No RAG2026.06 | 80.28 | |
| FL-RAGBackbone=Qwen2.5-7B, Paradigm=Naive RAG2026.06 | 80.28 | |
| IRCOTBackbone=Qwen2.5-7B, Paradigm=Agentic RAG2026.06 | 80.37 | |
| Mem1Backbone=Qwen2.5-3B, Paradigm=RL-based Agentic RAG2026.06 | 80.82 | |
| ARPOBackbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 82.2 | |
| FS-RAGBackbone=Qwen2.5-7B, Paradigm=Naive RAG2026.06 | 83.18 | |
| COTBackbone=Qwen2.5-3B, Paradigm=No RAG2026.06 | 83.47 | |
| BaseBackbone=Qwen2.5-3B, Paradigm=No RAG2026.06 | 85.73 | |
| AEPOBackbone=Qwen2.5-7B, Paradigm=RL-based Agentic RAG2026.06 | 90.07 |