Multi-hop Question Answering on HotpotQA (Exact Match, RAG Improvement)
52Exact Match AccuracyPyRAG
Evaluation results
| Method | Links | ||
|---|---|---|---|
| PyRAGBackbone=Qwen2.5-72B-Instruct2026.05 | 52 | 15.2 | |
| IRCoTBackbone=Qwen2.5-72B-Instruct2026.05 | 50.9 | — | |
| ITER-RETGENBackbone=Qwen2.5-72B-Instruct2026.05 | 50.5 | — | |
| Self-AskBackbone=Qwen2.5-72B-Instruct2026.05 | 48.2 | — | |
| Vanilla RAGBackbone=Qwen2.5-72B-Instruct2026.05 | 36.8 | — | |
| PyRAGBackbone=Qwen2.5-7B-Instruct2026.05 | 34 | 5.1 | |
| CoTBackbone=Qwen2.5-72B-Instruct2026.05 | 33.2 | — | |
| IRCoTBackbone=Qwen2.5-7B-Instruct2026.05 | 32.7 | — | |
| ITER-RETGENBackbone=Qwen2.5-7B-Instruct2026.05 | 32.5 | — | |
| Direct InferenceBackbone=Qwen2.5-72B-Instruct2026.05 | 30.6 | — | |
| Self-AskBackbone=Qwen2.5-7B-Instruct2026.05 | 30.2 | — | |
| Vanilla RAGBackbone=Qwen2.5-7B-Instruct2026.05 | 28.9 | — | |
| Direct InferenceBackbone=Qwen2.5-7B-Instruct2026.05 | 18.3 | — | |
| CoTBackbone=Qwen2.5-7B-Instruct2026.05 | 9.2 | — |