Long-context Question Answering on LongBench avg
45.4Avg. F1EfficientGraph-RAG
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| EfficientGraph-RAGParadigm=Hierarchical2026.05 | 45.4 | — | — | — | — | 36.2 | 708 | 1.87 | |
| ReActParadigm=Agentic2026.05 | 41.4 | — | — | — | — | 31.8 | 1,958 | 5.93 | |
| Vanilla RAGParadigm=Flat2026.05 | 40.4 | — | — | — | — | 31.3 | 798 | 0.6 | |
| Self-RAGParadigm=Agentic2026.05 | 39.6 | — | — | — | — | 30.7 | 1,935 | 6.05 | |
| LOUVER (OFFLOADED)Backbone=Llama-3.1-8B-Instruct, Memory Configuration=KV Offloading2026.05 | 38.9 | 0.07 | 6.02 | 285 | 28 | — | — | — | |
| Agentic-RAGParadigm=Agentic2026.05 | 35.2 | — | — | — | — | 24.8 | 2,466 | 5.8 | |
| InfLLMBackbone=Llama-3.1-8B-Instruct, Memory Configuration=KV Offloading2026.05 | 26.2 | 4.89 | 7.08 | 122 | 12 | — | — | — | |
| MagicPIGBackbone=Llama-3.1-8B-Instruct, Memory Configuration=KV Offloading2026.05 | 25.2 | 9.76 | 6.99 | 41 | 4 | — | — | — | |
| RetrievalAttentionBackbone=Llama-3.1-8B-Instruct, Memory Configuration=KV Offloading2026.05 | 25.1 | 53.95 | 7.05 | 0 | 0 | — | — | — | |
| GraphRAGParadigm=Graph2026.05 | 22.5 | — | — | — | — | 18.2 | 1,449 | 13.16 |