Multi-hop Question Answering on MuSiQue (EM, F1, Acc)
46.1F1C2RAG
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| C2RAGLLM=GPT-4o-mini2026.03 | 46.1 | 33.2 | — | — | |
| NeuroPathLLM=GPT-4o-mini2026.03 | 43.5 | 31.2 | — | — | |
| Iter-RetGenLLM=GPT-4o-mini2026.03 | 43.3 | 29.7 | — | — | |
| IRCoTLLM=GPT-4o-mini2026.03 | 43.1 | 31 | — | — | |
| CompactRAG GPT-4-ReadingToken / Sample=1.9K2026.02 | 42.34 | 30.8 | 43.6 | — | |
| HippoRAG2LLM=GPT-4o-mini2026.03 | 42.2 | 30.6 | — | — | |
| HippoRAGLLM=GPT-4o-mini2026.03 | 41.5 | 28.2 | — | — | |
| GFM-RAGLLM=GPT-4o-mini2026.03 | 40.9 | 30.5 | — | — | |
| CompactRAG LLaMA-3.1-8b-ReadingToken / Sample=1.9K2026.02 | 37.63 | 26.8 | 41.2 | — | |
| SPARC-RAGBackbone=Qwen3-32B2026.01 | 36.6 | 23.8 | 31.8 | 30.7 | |
| SubgraphRAGLLM=GPT-4o-mini2026.03 | 35.7 | 25.1 | — | — | |
| SPARC-RAG + DPOBackbone=Qwen3-32B, fine-tuned=true2026.01 | 33.9 | 21.4 | 28.8 | 28 | |
| G-retrieverLLM=GPT-4o-mini2026.03 | 33.9 | 23.1 | — | — | |
| Iter-RetGenToken / Sample=4.7K2026.02 | 32.42 | 24.8 | 40 | — | |
| SPARC-RAG + DPOBackbone=Qwen2.5-7B-Instruct, fine-tuned=true2026.01 | 32.4 | 20.8 | 25.2 | 26.1 | |
| DeepNoteBackbone=Qwen3-32B2026.01 | 30.9 | 18.6 | 30.8 | 26.8 | |
| BM25LLM=GPT-4o-mini2026.03 | 29.9 | 20.5 | — | — | |
| Iter-RetGenBackbone=Qwen3-32B2026.01 | 29.4 | 18.6 | 23.2 | 23.7 | |
| IRCoTToken / Sample=10.2K2026.02 | 29.08 | 21.2 | 32.4 | — | |
| Self-AskToken / Sample=6.9K2026.02 | 28.34 | 19.6 | 24.8 | — | |
| SPARC-RAGBackbone=Qwen2.5-7B-Instruct2026.01 | 24.3 | 16.2 | 17.8 | 19.4 | |
| PathRAGLLM=GPT-4o-mini2026.03 | 22.1 | 9.8 | — | — | |
| Iter-RetGenBackbone=Qwen2.5-7B-Instruct2026.01 | 21.4 | 13 | 19.8 | 18.1 | |
| Self-RAGBackbone=Qwen3-32B2026.01 | 20.7 | 10.4 | 12 | 14.4 | |
| DeepNoteBackbone=Qwen2.5-7B-Instruct2026.01 | 19.8 | 11.6 | 14.6 | 15.3 | |
| Speculative-RAGBackbone=Qwen3-32B2026.01 | 19.1 | 10.8 | 13.8 | 14.6 | |
| Self-RAGBackbone=Qwen2.5-7B-Instruct2026.01 | 15.8 | 8 | 11.6 | 11.8 | |
| Speculative-RAGBackbone=Qwen2.5-7B-Instruct2026.01 | 15.7 | 8 | 9.8 | 11.2 | |
| Vanilla RAGBackbone=Qwen2.5-7B-Instruct2026.01 | 13.8 | 6.6 | 11.4 | 10.6 | |
| Parallel (Naive)Backbone=Qwen2.5-7B-Instruct2026.01 | 13.8 | 6.4 | 11 | 10.4 | |
| IRCoTBackbone=Qwen2.5-7B-Instruct2026.01 | 12.4 | 4.8 | 11.8 | 9.7 | |
| Sequential (Naive)Backbone=Qwen2.5-7B-Instruct2026.01 | 12.1 | 5.2 | 10.8 | 9.4 | |
| Vanilla-RAGToken / Sample=2.7K2026.02 | 11.36 | 5.6 | 8.4 | — | |
| Parallel (Naive)Backbone=Qwen3-32B2026.01 | 10.9 | 2.6 | 14.8 | 9.4 | |
| Vanilla RAGBackbone=Qwen3-32B2026.01 | 10.6 | 2.4 | 15.8 | 9.6 | |
| Sequential (Naive)Backbone=Qwen3-32B2026.01 | 10.3 | 1.8 | 14.6 | 8.9 | |
| IRCoTBackbone=Qwen3-32B2026.01 | 10.1 | 5.4 | 6.6 | 7.4 | |
| LightRAGLLM=GPT-4o-mini2026.03 | 9.7 | 3.1 | — | — |