Multi-Hop Question Answering on HotpotQA (F1-score, Collapse Point)
63.09F1 ScoreCalibAdv
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CalibAdvBackbone=Qwen2.5-7B-Base2026.04 | 63.09 | — | |
| LLDBackbone=Qwen2.5-7B-Base2026.04 | 60.63 | — | |
| CalibAdvBackbone=Qwen2.5-3B-Base2026.04 | 59.19 | — | |
| CalibAdvBackbone=Llama3.2-3B-Instruct2026.04 | 56.97 | — | |
| SimpleTIRBackbone=Qwen2.5-3B-Base2026.04 | 56.54 | 310 | |
| Search-R1Backbone=Qwen2.5-3B-Base2026.04 | 56.52 | 292 | |
| LLDBackbone=Llama3.2-3B-Instruct2026.04 | 55.59 | — | |
| SimpleTIRBackbone=Qwen2.5-7B-Base2026.04 | 54.47 | 65 | |
| MT-GRPOBackbone=Qwen2.5-7B-Base2026.04 | 54 | — | |
| GiGPOBackbone=Qwen2.5-7B-Base2026.04 | 50.64 | — | |
| Search-R1Backbone=Qwen2.5-7B-Base2026.04 | 50.16 | 58 | |
| SimpleTIRBackbone=Llama3.2-3B-Instruct2026.04 | 49.6 | 54 | |
| StepSearchBackbone=Qwen2.5-7B-Base2026.04 | 48.84 | — | |
| Search-R1Backbone=Llama3.2-3B-Instruct2026.04 | 48.08 | 40 |