Multi-hop Question Answering on HotpotQA (EM, Tok)
47.9EMCARL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CARLModel Scale=7B-Inst.2026.05 | 47.9 | 1,193 | |
| Search-R1 PPOModel Scale=7B-Inst.2026.05 | 42.8 | 1,837 | |
| CARLModel Scale=3B-Inst.2026.05 | 40.4 | 1,324 | |
| Search-R1 GRPOModel Scale=7B-Inst.2026.05 | 39.1 | 1,779 | |
| Search-R1 GRPOModel Scale=3B-Inst.2026.05 | 34.3 | 1,921 | |
| SFT (rej. sampled)Model Scale=7B-Inst.2026.05 | 33.7 | 1,712 | |
| CoT+Tools (selective)Model Scale=7B-Inst.2026.05 | 31.1 | 1,621 | |
| CoT+Tools (always)Model Scale=7B-Inst.2026.05 | 30.8 | 1,862 | |
| Search-R1 PPOModel Scale=3B-Inst.2026.05 | 30.2 | 1,809 | |
| SFT (rej. sampled)Model Scale=3B-Inst.2026.05 | 24.6 | 1,818 | |
| CoT+Tools (always)Model Scale=3B-Inst.2026.05 | 22.2 | 1,943 | |
| CoT+Tools (selective)Model Scale=3B-Inst.2026.05 | 19.1 | 1,685 |