Search-Augmented Multi-Turn QA on HotpotQA (EM/F1)
40.2Exact Match (EM)IGPO (gold-sup.)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| IGPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 40.2 | 49.9 | |
| EMPOBackbone=Qwen3-8B2026.05 | 39.2 | 44.5 | |
| GRPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 39.1 | 48.4 | |
| TTRLBackbone=Qwen3-8B2026.05 | 38.5 | 43.1 | |
| IGPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 35.9 | 45.1 | |
| SIOPBackbone=Qwen3-8B2026.05 | 34.8 | 45.3 | |
| SIOPBackbone=Qwen3-4B2026.05 | 34.4 | 44.3 | |
| GRPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 33 | 44.1 | |
| EMPOBackbone=Qwen3-4B2026.05 | 33 | 41.4 | |
| TTRLBackbone=Qwen3-4B2026.05 | 31.8 | 40.1 |