Search-augmented multi-turn question answering on TriviaQA
64.6Exact Match (EM) AccuracySIOP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SIOPBackbone=Qwen3-8B2026.05 | 64.6 | 73.6 | |
| GRPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 64.3 | 72.8 | |
| IGPO (gold-sup.)Backbone=Qwen3-8B2026.05 | 64.1 | 73 | |
| EMPOBackbone=Qwen3-8B2026.05 | 63 | 71.7 | |
| GRPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 62.3 | 70.3 | |
| TTRLBackbone=Qwen3-8B2026.05 | 62.3 | 72.2 | |
| EMPOBackbone=Qwen3-4B2026.05 | 61.1 | 69.6 | |
| TTRLBackbone=Qwen3-4B2026.05 | 60.7 | 69.1 | |
| SIOPBackbone=Qwen3-4B2026.05 | 60.4 | 69.1 | |
| IGPO (gold-sup.)Backbone=Qwen3-4B2026.05 | 59.8 | 68.2 |