Multi-hop QA on 2WikiMultihopQA (pass@1 accuracy)
81.5Pass@1 AccuracySwiR
Evaluation Results
| Method | Links | |
|---|---|---|
| SwiRBackbone=Qwen3-8B2025.10 | 81.5 | |
| CoT (Greedy)Backbone=Qwen3-8B, decoding=Greedy2025.10 | 79.5 | |
| CoTBackbone=Qwen3-8B2025.10 | 79 | |
| Soft ThinkingBackbone=Qwen3-8B2025.10 | 79 |