Logical Reasoning on LogiQA (Acc, Len, Faith)
68.9Accuracy (LogiQA)GeoFaith
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GeoFaithBackbone=Qwen3-4B2026.05 | 68.9 | 1,000 | 90.2 | |
| THSBackbone=Qwen3-4B2026.05 | 68.2 | 1,000 | 84.6 | |
| GRPOBackbone=Qwen3-4B2026.05 | 65.9 | 1,400 | 87.1 | |
| KnowRLBackbone=Qwen3-4B2026.05 | 65.3 | 1,200 | 88 | |
| OriginalBackbone=Qwen3-4B2026.05 | 65.1 | 2,300 | 88.7 | |
| GeoFaithBackbone=Qwen3-1.7B2026.05 | 52.4 | 300 | 46.2 | |
| TruthRLBackbone=Qwen3-4B2026.05 | 52.2 | 600 | 74 | |
| GRPOBackbone=Qwen3-1.7B2026.05 | 51.2 | 400 | 46.3 | |
| KnowRLBackbone=Qwen3-1.7B2026.05 | 47.9 | 500 | 42.8 | |
| THSBackbone=Qwen3-1.7B2026.05 | 46.4 | 200 | 33.2 | |
| TruthRLBackbone=Qwen3-1.7B2026.05 | 46.2 | 300 | 19.2 | |
| OriginalBackbone=Qwen3-1.7B2026.05 | 34 | 600 | 45.7 |