Multi-hop Question Answering on MuSiQue (Recall@K, NDCG@K)
36.8Recall@1F-GRPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| F-GRPOBackbone=Qwen3-4B2026.05 | 36.8 | 71.3 | 71.8 | 91.6 | 78.5 | 76.5 | |
| GRPOBackbone=Qwen3-4B2026.05 | 36.4 | 63 | 63 | 90.8 | 71.3 | 69.2 | |
| F-GRPOBackbone=Qwen3.5-2B2026.05 | 36.2 | 72.7 | 73.2 | 89.9 | 79.2 | 77.2 | |
| GRPOBackbone=Qwen3.5-2B2026.05 | 36.1 | 63.1 | 65.2 | 90 | 71.4 | 70.6 | |
| MonoT52026.05 | 34.9 | 64.5 | 74.9 | 86.1 | 70.7 | 74.6 | |
| DuoT52026.05 | 34.3 | 62.8 | 73.5 | 84.8 | 69.2 | 73.3 | |
| LiT52026.05 | 33.6 | 54.7 | 64.4 | 82.7 | 61.9 | 65.6 | |
| RankZephyr2026.05 | 33.5 | 61.3 | 71.5 | 82.3 | 67.1 | 71.1 | |
| Decoupled SFT (rank. only)Backbone=Qwen3-4B2026.05 | 32.3 | 63 | 68.3 | 79.8 | 68.6 | 69.7 | |
| SFTBackbone=Qwen3-4B2026.05 | 27.9 | 54.4 | 60.9 | 69.5 | 59.5 | 61.5 | |
| Decoupled SFT (full)Backbone=Qwen3-4B2026.05 | 24 | 39 | 44.2 | 60.6 | 44.9 | 46.5 | |
| Zero-shot (single-step)Backbone=Qwen3-4B2026.05 | 23.4 | 48.1 | 56.3 | 58.4 | 51.6 | 54.7 | |
| BM252026.05 | 22.9 | 42.9 | 53.4 | 57 | 46.7 | 51.4 | |
| Decoupled SFT (rank. only)Backbone=Qwen3.5-2B2026.05 | 17.8 | 37.4 | 49.8 | 43.7 | 39 | 44.8 | |
| SFTBackbone=Qwen3.5-2B2026.05 | 13.5 | 29.2 | 38 | 34.4 | 31.2 | 35 | |
| Decoupled SFT (sel. only)Backbone=Qwen3-4B2026.05 | 8.6 | 16.3 | 18.3 | 21.7 | 18 | 18.5 | |
| Zero-shot (single-step)Backbone=Qwen3.5-2B2026.05 | 5.2 | 10.6 | 12.4 | 13.2 | 11.5 | 12.1 | |
| Decoupled SFT (full)Backbone=Qwen3.5-2B2026.05 | 5.2 | 14.8 | 24.5 | 13.8 | 14.7 | 19.5 | |
| Random2026.05 | 5 | 15 | 24.7 | 13.3 | 14.9 | 19.6 | |
| Zero-shot (two-step)Backbone=Qwen3-4B2026.05 | 4.1 | 8.5 | 9 | 10.2 | 9.1 | 9.2 | |
| Zero-shot (two-step)Backbone=Qwen3.5-2B2026.05 | 2.2 | 4.1 | 4.8 | 5.4 | 4.5 | 4.7 | |
| Decoupled SFT (sel. only)Backbone=Qwen3.5-2B2026.05 | 0.3 | 0.6 | 0.7 | 0.7 | 0.6 | 0.7 |