Multi-hop Reasoning on 2WikiMultihopQA (Acc, Len, Faith)
82.1AccuracyGeoFaith
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GeoFaithBackbone=Qwen3-4B2026.05 | 82.1 | 0.3 | 92 | |
| GRPOBackbone=Qwen3-4B2026.05 | 81.5 | 0.4 | 84.8 | |
| KnowRLBackbone=Qwen3-4B2026.05 | 79.8 | 0.4 | 86.5 | |
| OriginalBackbone=Qwen3-4B2026.05 | 79.2 | 0.6 | 90.1 | |
| TruthRLBackbone=Qwen3-4B2026.05 | 78.1 | 0.2 | 79.5 | |
| THSBackbone=Qwen3-4B2026.05 | 76.6 | 0.3 | 86.8 | |
| GeoFaithBackbone=Qwen3-1.7B2026.05 | 71.2 | 0.3 | 77.5 | |
| GRPOBackbone=Qwen3-1.7B2026.05 | 67.4 | 0.6 | 71.9 | |
| THSBackbone=Qwen3-1.7B2026.05 | 56.5 | 0.2 | 49.3 | |
| TruthRLBackbone=Qwen3-1.7B2026.05 | 53.6 | 0.2 | 53.9 | |
| KnowRLBackbone=Qwen3-1.7B2026.05 | 50.5 | 0.2 | 60.3 | |
| OriginalBackbone=Qwen3-1.7B2026.05 | 34.5 | 0.3 | 52 |