Multi-Hop Question Answering on Bamboogle (F1, Collapse Point)
62.25F1 ScoreCalibAdv
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CalibAdvBackbone=Qwen2.5-7B-Base2026.04 | 62.25 | — | |
| LLDBackbone=Qwen2.5-7B-Base2026.04 | 58.36 | — | |
| StepSearchBackbone=Qwen2.5-7B-Base2026.04 | 55.74 | — | |
| CalibAdvBackbone=Llama3.2-3B-Instruct2026.04 | 54.2 | — | |
| LLDBackbone=Llama3.2-3B-Instruct2026.04 | 53.34 | — | |
| MT-GRPOBackbone=Qwen2.5-7B-Base2026.04 | 51.47 | — | |
| SimpleTIRBackbone=Qwen2.5-7B-Base2026.04 | 50.9 | 65 | |
| CalibAdvBackbone=Qwen2.5-3B-Base2026.04 | 50.62 | — | |
| Search-R1Backbone=Qwen2.5-3B-Base2026.04 | 50.48 | 292 | |
| Search-R1Backbone=Qwen2.5-7B-Base2026.04 | 48.66 | 58 | |
| SimpleTIRBackbone=Qwen2.5-3B-Base2026.04 | 48.58 | 310 | |
| Search-R1Backbone=Llama3.2-3B-Instruct2026.04 | 47.33 | 40 | |
| GiGPOBackbone=Qwen2.5-7B-Base2026.04 | 47.11 | — | |
| SimpleTIRBackbone=Llama3.2-3B-Instruct2026.04 | 45.93 | 54 |