Multi-hop Question Answering on 2WikiMultihopQA OOD (Exact Match)
39.2Exact Match AccuracyRECON
Evaluation Results
| Method | Links | |
|---|---|---|
| RECONBackbone=Qwen2.5-7B-Base, Training Protocol=PPO, RECON Stage=Stage 1 + Stage 22025.10 | 39.2 | |
| Search-R1Backbone=Qwen2.5-7B-Base, Training Protocol=PPO2025.10 | 38.2 | |
| RECONBackbone=Qwen2.5-3B-Base, Training Protocol=PPO, RECON Stage=Stage 1 + Stage 22025.10 | 31.9 | |
| Rejection Sampling2025.10 | 29.6 | |
| R1-instruct2025.10 | 29.2 | |
| RECONBackbone=Qwen2.5-7B-Base, Training Protocol=PPO, RECON Stage=Stage 2 Distillation-only2025.10 | 28.6 | |
| R1-base2025.10 | 27.3 | |
| Search-R1Backbone=Qwen2.5-3B-Base, Training Protocol=PPO2025.10 | 27.3 | |
| SFT2025.10 | 25.9 | |
| Direct Inference2025.10 | 25 | |
| RAG2025.10 | 23.5 | |
| RECONBackbone=Qwen2.5-3B-Base, Training Protocol=PPO, RECON Stage=Stage 2 Distillation-only2025.10 | 23 | |
| Search-o12025.10 | 17.6 | |
| IRCoT2025.10 | 14.9 | |
| CoT2025.10 | 11.1 |