Multi-Hop Question Answering on HotpotQA, 2Wiki, Musique, and Bamboogle
64.6HotpotQA ScoreQwen2.5-32B-Instruct + Proposed Method
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Qwen2.5-32B-Instruct + Proposed MethodBackbone=Qwen2.5-32B-Instruct, Training=Proposed Method2026.05 | 64.6 | 55 | 31.6 | 60 | — | 58.5 | |
| Qwen2.5-14B-Instruct + Proposed MethodBackbone=Qwen2.5-14B-Instruct, Training=Proposed Method2026.05 | 61 | 50.8 | 29.8 | 58.4 | — | 56.1 | |
| SLEA-RLBase Model=Qwen2.5-7B-Instruct, Type=RL Methods, Training Context=Trained on NQ and HotpotQA2026.03 | 59.8 | 70.5 | 77.2 | 33.2 | — | 60.9 | |
| Qwen2.5-14B-Instruct + SSPBackbone=Qwen2.5-14B-Instruct, Training=SSP2026.05 | 58 | 46.8 | 27.2 | 58.4 | — | 54.2 | |
| TRACEModel=Qwen3-14B2026.06 | 57.8 | 53.1 | 33.7 | 71.3 | — | 54 | |
| Qwen2.5-32B-Instruct + SSPBackbone=Qwen2.5-32B-Instruct, Training=SSP2026.05 | 57.4 | 47.2 | 30.8 | 61 | — | 55.2 | |
| IGPOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 57.2 | 68.2 | 31.4 | 74.9 | — | 58.7 | |
| TRACEModel=Qwen3-8B2026.06 | 55.7 | 50.9 | 29.5 | 66.3 | — | 50.6 | |
| PCLModel=Qwen3-14B2026.06 | 55.6 | 52.9 | 30.2 | 67.2 | — | 51.5 | |
| GRPOModel=Qwen3-14B2026.06 | 55.3 | 52.7 | 30.1 | 66.8 | — | 51.2 | |
| TreePOModel=Qwen3-14B2026.06 | 55.1 | 55.3 | 31 | 70.6 | — | 53 | |
| TreePOModel=Qwen3-8B2026.06 | 54 | 49.7 | 28.4 | 65.9 | — | 49.5 | |
| Qwen2.5-14B-InstructBackbone=Qwen2.5-14B-Instruct2026.05 | 53.6 | 44.8 | 23.8 | 59.2 | — | 52.3 | |
| Qwen2.5-32B-InstructBackbone=Qwen2.5-32B-Instruct2026.05 | 53.6 | 46.6 | 24.2 | 53.6 | — | 52.4 | |
| PCLModel=Qwen3-8B2026.06 | 53.6 | 49.2 | 27.9 | 65 | — | 48.9 | |
| GRPOModel=Qwen3-8B2026.06 | 53.2 | 48.7 | 27.4 | 64.7 | — | 48.5 | |
| RLOOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 49.6 | 55 | 62.2 | 24.8 | — | 49.7 | |
| Search-R1-GRPO + LLDSBackbone=Qwen2.5-7b-Ins2025.12 | 49.2 | 49 | 22.8 | 49.2 | 42.6 | 48.3 | |
| GRPOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 48.9 | 57.7 | 65.1 | 25 | — | 51.9 | |
| Search-R1-GRPO + LLDSBackbone=Qwen2.5-7b-Base2025.12 | 48.6 | 46.9 | 23.2 | 50.8 | 42.4 | 48 | |
| PPOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 48.6 | 59.7 | 63.4 | 26.2 | — | 51.5 | |
| ReActModel=Qwen3-14B2026.06 | 46.4 | 44.2 | 22.8 | 55.3 | — | 42.2 | |
| GSPOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 46.3 | 60.1 | 67.6 | 25.4 | — | 52 | |
| Reinforce++Base Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 45.9 | 54.5 | 61.2 | 23.7 | — | 47.3 | |
| SkillRLBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 45.9 | 43.2 | 20.2 | 40.3 | — | 47.1 | |
| GRPO-LLDS-MABackbone=Llama3.2-3b-Base2025.12 | 44.9 | 46.2 | 16.4 | 29.8 | 34.3 | 42.7 | |
| Tree-GRPOBackbone=Qwen2.5-7B2025.12 | 44.6 | 42.3 | 20.2 | 44 | 37.8 | 43.9 | |
| CriticSearchBackbone=Qwen2.5-7B2025.12 | 44.2 | 42.8 | 19.4 | 47.2 | 38.4 | — | |
| Search-R1-PPO-BaseBackbone=Qwen2.5-7B-Base2026.02 | 43.3 | 38.2 | 19.6 | 43.2 | 36.1 | 43.1 | |
| Search-R1-PPO-BaseBackbone=Qwen2.5-7B, Alignment=Base2025.12 | 43.3 | 38.2 | 19.6 | 43.2 | 36.1 | 43.1 | |
| DARTBackbone=Qwen2.5-7B-Instruct2026.02 | 43.1 | 34.9 | 16.3 | 38.6 | 33 | 42 | |
| DARTBackbone=Qwen2.5-7B-Base2026.02 | 42.5 | 33.8 | 15.5 | 37.6 | 32.3 | 41.6 | |
| ReActModel=Qwen3-8B2026.06 | 41.9 | 37.6 | 18.4 | 44.7 | — | 35.7 | |
| GiGPOBase Model=Qwen2.5-7B-Instruct, Type=RL Methods2026.03 | 41.6 | 43.6 | 18.9 | 68.9 | — | 47.2 | |
| Search-R1-GRPOBackbone=Qwen2.5-7B-Instruct2026.02 | 38.6 | 34.6 | 16.2 | 40 | 32.4 | 39.6 | |
| Search-R1-GRPOBackbone=Qwen2.5-7b-Ins2025.12 | 38.6 | 34.6 | 16.2 | 40 | 32.4 | 39.6 | |
| Search-R1-PPO-InsBackbone=Qwen2.5-7B-Instruct2026.02 | 37 | 41.4 | 14.6 | 36.8 | 32.5 | 38.5 | |
| Search-R1-PPO-InsBackbone=Qwen2.5-7B, Alignment=Instruct2025.12 | 37 | 41.4 | 14.6 | 36.8 | 32.5 | 38.5 | |
| GRPO-LLDSBackbone=Llama3.2-3b-Base2025.12 | 36.6 | 29.8 | 11.3 | 24.2 | 25.5 | 37.4 | |
| ZeroSearch-InsBackbone=Qwen2.5-7B, Alignment=Instruct2025.12 | 34.6 | 35.2 | 18.4 | 27.8 | 29 | 39.1 | |
| StepSearchBackbone=Qwen2.5-7B2025.12 | 34.6 | 39.4 | 18.3 | 44.4 | 34.2 | 38.2 | |
| Rejection SamplingBackbone=Qwen2.5-7B2026.02 | 33.1 | 29.6 | 12.3 | 35.5 | 27.6 | 34.8 | |
| Rejection SamplingBackbone=Qwen2.5-7B2025.12 | 33.1 | 29.6 | 12.3 | 35.5 | 27.6 | 34.8 | |
| Search-R1-GRPOBackbone=Qwen2.5-7B-Base2026.02 | 32.6 | 29.7 | 12.5 | 36 | 27.7 | 35 | |
| Search-R1-GRPOBackbone=Qwen2.5-7b-Base2025.12 | 32.6 | 29.7 | 12.5 | 36 | 27.7 | 35 | |
| Search-R1++Base Model=Qwen2.5-3B, Method Category=RL trained Deep Research Agents2026.02 | 32.5 | 30 | 6.5 | 16.2 | — | — | |
| ZeroSearch-BaseBackbone=Qwen2.5-7B, Alignment=Base2025.12 | 32 | 34 | 18 | 33.3 | 29.3 | 40.9 | |
| RAGBackbone=Qwen2.5-7B2026.02 | 29.9 | 23.5 | 5.8 | 20.8 | 20 | 30.4 | |
| RAGBackbone=Qwen2.5-7B2025.12 | 29.9 | 23.5 | 5.8 | 20.8 | 20 | 30.4 | |
| Search-R1Base Model=Qwen2.5-3B, Method Category=RL trained Deep Research Agents2026.02 | 26.3 | 25.4 | 4.8 | 10.9 | — | — | |
| R1-baseBackbone=Qwen2.5-7B-Base2026.02 | 24.2 | 27.3 | 8.3 | 20.3 | 20 | 26.2 | |
| R1-baseBackbone=Qwen2.5-7B, Alignment=Base2025.12 | 24.2 | 27.3 | 8.3 | 20.3 | 20 | 26.2 | |
| R1-instructBackbone=Qwen2.5-7B-Instruct2026.02 | 23.7 | 29.2 | 7.2 | 29.3 | 22.4 | 27.1 | |
| R1-instructBackbone=Qwen2.5-7B, Alignment=Instruct2025.12 | 23.7 | 29.2 | 7.2 | 29.3 | 22.4 | 27.1 | |
| SFTBackbone=Qwen2.5-7B2026.02 | 21.7 | 25.9 | 6.6 | 11.2 | 16.4 | 20.7 | |
| R1-baseBase Model=Qwen2.5-3B, Method Category=RL Trained LLM without Retrieval2026.02 | 20.1 | 26.8 | 5.5 | 22.4 | — | — | |
| GRPOBackbone=Llama3.2-3b-Base2025.12 | 19.6 | 23.2 | 4.4 | 9.7 | 14.2 | 21.2 | |
| Search-o1Backbone=Qwen2.5-7B2026.02 | 18.7 | 17.6 | 6.2 | 29.6 | 18 | 20.6 | |
| Search-o1Backbone=Qwen2.5-7B2025.12 | 18.7 | 17.6 | 6.2 | 29.6 | 18 | 20.6 | |
| Direct InferenceBackbone=Qwen2.5-7B2026.02 | 18.3 | 25 | 3.1 | 12 | 14.6 | 18.1 | |
| Direct InferenceBackbone=Qwen2.5-7B2025.12 | 18.3 | 25 | 3.1 | 12 | 14.6 | 18.1 | |
| IRCoTBackbone=Qwen2.5-7B2026.02 | 13.3 | 14.9 | 7.2 | 22.4 | 14.5 | 23.9 | |
| CoTBackbone=Qwen2.5-7B2026.02 | 9.2 | 11.1 | 2.2 | 23.2 | 11.4 | 10.6 | |
| ReActBase Model=Qwen2.5-3B, Method Category=Training free Deep Research Agent2026.02 | 4.7 | 4.6 | 0.8 | 1.6 | — | — |