Fact Verification on FEVER held-out (test)
95.1AccuracyDirect
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DirectModel=Qwen3-14B2026.06 | 95.1 | 55.3 | — | — | — | |
| Design BModel=Qwen3-14B2026.06 | 94.7 | 79.6 | 0.5 | 63.4 | 0.4 | |
| Design A rescueModel=Qwen3-14B2026.06 | 94.4 | 79.3 | 0.2 | 63.3 | 0.3 | |
| Design A 2-sidedModel=Qwen3-14B2026.06 | 94.4 | 79.3 | 0.2 | 63.3 | 0.3 | |
| SIFTModel=Qwen3-14B2026.06 | 94 | 79.1 | — | 63 | — | |
| NLI-as-judge (DeBERTa)2026.06 | 91.7 | — | — | — | — | |
| Design A rescueModel=Qwen3-4B2026.06 | 88.8 | 75.1 | -0.4 | 61.5 | -1.1 | |
| Design A 2-sidedModel=Qwen3-4B2026.06 | 88.8 | 75.1 | -0.4 | 61.5 | -1.1 | |
| Design BModel=Qwen3-4B2026.06 | 88.8 | 75.1 | -0.4 | 61.5 | -1 | |
| DirectModel=Qwen3-4B2026.06 | 88.5 | 71.5 | — | 61.1 | — | |
| DirectModel=Gemma2-9B2026.06 | 88.1 | 49.6 | — | 41 | — | |
| SIFTModel=Qwen3-4B2026.06 | 85.3 | 75.5 | — | 62.5 | — | |
| Design BModel=Llama3.1-8B2026.06 | 83 | 67 | 2.6 | 54.6 | 1.6 | |
| DirectModel=Llama3.1-8B2026.06 | 82.3 | 50.8 | — | 45.5 | — | |
| Design A rescueModel=Llama3.1-8B2026.06 | 79.2 | 64.6 | 0.2 | 53 | 0 | |
| Design A 2-sidedModel=Llama3.1-8B2026.06 | 79.2 | 64.6 | 0.2 | 53 | 0 | |
| SIFTModel=Gemma2-9B2026.06 | 78.2 | 60.2 | — | 48.3 | — | |
| Design A rescueModel=Gemma2-9B2026.06 | 78.2 | 60.2 | 0 | 48.3 | 0 | |
| Design A 2-sidedModel=Gemma2-9B2026.06 | 78.2 | 60.2 | 0 | 48.3 | 0 | |
| Design BModel=Gemma2-9B2026.06 | 78.2 | 60.2 | 0 | 48.3 | 0 | |
| SIFTModel=Llama3.1-8B2026.06 | 74.9 | 64.4 | — | 53 | — |