Search-augmented multi-turn Question Answering on Bamboogle (EM, F1)
48.8Exact Match (EM)GRPO (gold-sup.)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 48.8 | 59.9 | |
| SIOPBackbone=Qwen3-8B2026.05 | 48.8 | 61.2 | |
| EMPOBackbone=Qwen3-8B2026.05 | 47.2 | 58 | |
| GRPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 46.4 | 53.4 | |
| IGPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 46.4 | 55 | |
| EMPOBackbone=Qwen3-4B2026.05 | 45.6 | 51 | |
| IGPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 45.6 | 58.3 | |
| TTRLBackbone=Qwen3-8B2026.05 | 44 | 53.9 | |
| SIOPBackbone=Qwen3-4B2026.05 | 40.8 | 52.8 | |
| TTRLBackbone=Qwen3-4B2026.05 | 40 | 52 |