Open-domain Question Answering on Natural Questions (NQ) (test)
58.15AccuracyClean
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CleanLLM Backbone=Qwen3-Max, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 58.15 | — | — | |
| PoisonedRAG-BBLLM Backbone=Qwen3-Max, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 58.07 | — | 12.33 | |
| Joint-GCGLLM Backbone=Qwen3-Max, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 57.06 | — | 31.82 | |
| The RAG ParadoxLLM Backbone=Qwen3-Max, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 56.13 | — | 23.04 | |
| INFO-RAGBackbone=LLaMA-2-13B-chat, Zero-shot=true2024.02 | 54.04 | 46.55 | — | |
| INFO-RAGBackbone=LLaMA-2-7B-chat, Zero-shot=true2024.02 | 53.93 | 43.78 | — | |
| LLaMA-2-13B-chatZero-shot=true2024.02 | 50.36 | 43.23 | — | |
| LLaMA-2-7B-chatZero-shot=true2024.02 | 49.42 | 40.83 | — | |
| INFO-RAGBackbone=LLaMA-2-13B, Zero-shot=true2024.02 | 47.82 | 41.04 | — | |
| LLaMA-2-13BZero-shot=true2024.02 | 47.4 | 36.86 | — | |
| LLaMA-2-7BZero-shot=true2024.02 | 46.82 | 35.78 | — | |
| CleanLLM Backbone=DeepSeek-R1, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 46.33 | — | — | |
| INFO-RAGBackbone=LLaMA-2-7B, Zero-shot=true2024.02 | 45.74 | 39.83 | — | |
| CleanLLM Backbone=GPT-4o, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 45.09 | — | — | |
| The RAG ParadoxLLM Backbone=DeepSeek-R1, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 45.07 | — | 23.98 | |
| Joint-GCGLLM Backbone=DeepSeek-R1, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 44.15 | — | 27.71 | |
| PoisonedRAG-BBLLM Backbone=DeepSeek-R1, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 43.86 | — | 12.05 | |
| Joint-GCGLLM Backbone=Qwen2.5-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 42.3 | — | 33.07 | |
| The RAG ParadoxLLM Backbone=GPT-4o, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 42.15 | — | 15.13 | |
| DPR-multiCategory=Retriever-Reader, Sparse=false, Storage (GB)=76, #Q/sec (GPU)=0.9, #Q/sec (CPU)=0.042020.12 | 41.5 | — | — | |
| CleanLLM Backbone=Vicuna-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 41.13 | — | — | |
| PoisonedRAG-BBLLM Backbone=Qwen2.5-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 41.1 | — | 14.81 | |
| DensePhrasesCategory=Phrase Retrieval, Sparse=false, Storage (GB)=320, #Q/sec (GPU)=20.6, #Q/sec (CPU)=13.62020.12 | 40.9 | — | — | |
| CleanLLM Backbone=Qwen2.5-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 40.89 | — | — | |
| REALMNewsCategory=Retriever-Reader, Sparse=false, Storage (GB)=18, #Q/sec (GPU)=8.4, #Q/sec (CPU)=1.22020.12 | 40.4 | — | — | |
| PoisonedRAG-BBLLM Backbone=GPT-4o, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 40.27 | — | 13.25 | |
| The RAG ParadoxLLM Backbone=Qwen2.5-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 40.25 | — | 14.17 | |
| The RAG ParadoxLLM Backbone=Vicuna-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 40.09 | — | 15.06 | |
| PoisonedRAG-BBLLM Backbone=Vicuna-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 39.97 | — | 13.79 | |
| Joint-GCGLLM Backbone=Vicuna-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 39.75 | — | 33.29 | |
| PoisonedRAG-BBLLM Backbone=LLaMA-2-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 38.16 | — | 13.64 | |
| Joint-GCGLLM Backbone=GPT-4o, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 37.98 | — | 32.16 | |
| CleanLLM Backbone=LLaMA-2-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 37.5 | — | — | |
| The RAG ParadoxLLM Backbone=LLaMA-2-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 36.7 | — | 13.89 | |
| Joint-GCGLLM Backbone=LLaMA-2-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 35.03 | — | 33.18 | |
| ORQACategory=Retriever-Reader, Sparse=false, Storage (GB)=18, #Q/sec (GPU)=8.6, #Q/sec (CPU)=1.22020.12 | 33.3 | — | — | |
| CRCPLLM Backbone=LLaMA-2-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 15.78 | — | 93.09 | |
| CRCPLLM Backbone=Qwen3-Max, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 15.2 | — | 94.07 | |
| CRCPLLM Backbone=DeepSeek-R1, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 15.18 | — | 95.88 | |
| CRCPLLM Backbone=Vicuna-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 14.92 | — | 94.76 | |
| DenSPI + SparcCategory=Phrase Retrieval, Sparse=true, Storage (GB)=1547, #Q/sec (GPU)=2.1, #Q/sec (CPU)=1.72020.12 | 14.5 | — | — | |
| CRCPLLM Backbone=Qwen2.5-7B, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 14.16 | — | 91.53 | |
| CRCPLLM Backbone=GPT-4o, Pipeline=Chunking, Retrieval, and Reranking2026.06 | 13.59 | — | 92.19 | |
| DenSPICategory=Phrase Retrieval, Sparse=true, Storage (GB)=1200, #Q/sec (GPU)=2.9, #Q/sec (CPU)=2.42020.12 | 8.1 | — | — |