Multi-hop Question Answering on HotpotQA (EM, F1, Token, LM Lat.)
51.6EMReAct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ReActParadigm=Agentic2026.05 | 51.6 | 64.5 | 1,678 | 5.94 | |
| EfficientGraph-RAGParadigm=Hierarchical2026.05 | 51.1 | 64 | 478 | 1.94 | |
| Self-RAGParadigm=Agentic2026.05 | 48.7 | 61.7 | 1,391 | 5.39 | |
| Agentic-RAGParadigm=Agentic2026.05 | 48.7 | 63 | 1,594 | 4.31 | |
| Vanilla RAGParadigm=Flat2026.05 | 48 | 60.3 | 586 | 0.45 | |
| VanillaModel=Qwen3-4B-Thinking-2507, Calibration Threshold Setup=calibrated on Bespoke-Stratos-17k without recalibration2026.06 | 34.95 | — | 3,511 | — | |
| GraphRAGParadigm=Graph2026.05 | 34.2 | 43.9 | 7,177 | 70.73 | |
| CUSUMModel=Qwen3-4B-Thinking-2507, Calibration Threshold Setup=calibrated on Bespoke-Stratos-17k without recalibration2026.06 | 34.1 | — | 3,080 | — | |
| DEERModel=Qwen3-4B-Thinking-2507, Calibration Threshold Setup=calibrated on Bespoke-Stratos-17k without recalibration2026.06 | 33.2 | — | 3,392 | — | |
| DynasorModel=Qwen3-4B-Thinking-2507, Calibration Threshold Setup=calibrated on Bespoke-Stratos-17k without recalibration2026.06 | 23.2 | — | 2,378 | — | |
| Baseline2026.05 | — | 46.64 | — | — | |
| DenseSteer2026.05 | — | 45.87 | — | — | |
| InFamilySteer2026.05 | — | 47.62 | — | — |