Multi-Hop Reasoning on MuSiQue (Accuracy and Search Efficiency)
10.4AccuracyBES
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BESBackbone=Llama-3.1-8B-Instruct2026.05 | 10.4 | 2.11 | 3.05 | 94 | |
| Tree-GRPOBackbone=Llama-3.1-8B-Instruct2026.05 | 7.4 | 0.65 | 1.36 | 71 | |
| BESBackbone=Llama-3.2-3B-Instruct2026.05 | 7 | 2.31 | 3.29 | 97 | |
| Base modelBackbone=Llama-3.1-8B-Instruct2026.05 | 6.6 | — | — | — | |
| GRPOBackbone=Llama-3.1-8B-Instruct2026.05 | 5.6 | 1.46 | 1.83 | 37 | |
| Base modelBackbone=Llama-3.2-3B-Instruct2026.05 | 4 | — | — | — | |
| Tree-GRPOBackbone=Llama-3.2-3B-Instruct2026.05 | 3.9 | 1.5 | 2.14 | 64 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.05 | 2.1 | 0.84 | 0.2 | 64 |