Search-augmented multi-turn Question Answering on Natural Questions (NQ)
28.3Exact Match (EM) AccuracyGRPO (gold-sup.)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 28.3 | 40.8 | |
| SIOPBackbone=Qwen3-4B2026.05 | 28.1 | 40 | |
| IGPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 27.5 | 40.7 | |
| EMPOBackbone=Qwen3-4B2026.05 | 27 | 40.4 | |
| GRPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 27 | 42.2 | |
| SIOPBackbone=Qwen3-8B2026.05 | 27 | 40.3 | |
| TTRLBackbone=Qwen3-8B2026.05 | 26.6 | 40.1 | |
| IGPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 26.2 | 39.6 | |
| TTRLBackbone=Qwen3-4B2026.05 | 26.2 | 40.2 | |
| EMPOBackbone=Qwen3-8B2026.05 | 25.1 | 40.6 |