Question Answering on TruthfulQA o=1 Semantic-level
90.9AccuracyW/O Decontamination
Evaluation results
| Method | Links | ||
|---|---|---|---|
| W/O DecontaminationModel=LLaMA-3, Split=fcon, Access=-2026.01 | 90.9 | 0.442 | |
| W/O DecontaminationModel=Mistral, Split=fcon, Access=-2026.01 | 89 | 0.641 | |
| ITDModel=Mistral, Split=fcon, Access=Black-box2026.01 | 89 | 0.641 | |
| Short CircuitModel=Mistral, Split=fcon, Access=White-box2026.01 | 88 | 0.631 | |
| TEDModel=LLaMA-3, Split=fcon, Access=Black-box2026.01 | 87.6 | 0.409 | |
| ITDModel=LLaMA-3, Split=fcon, Access=Black-box2026.01 | 87.6 | 0.409 | |
| W/O DecontaminationModel=Qwen 2.5, Split=fcon, Access=-2026.01 | 85.6 | 0.325 | |
| ITDModel=Qwen 2.5, Split=fcon, Access=Black-box2026.01 | 85.6 | 0.325 | |
| TEDModel=Qwen 2.5, Split=fcon, Access=Black-box2026.01 | 81.3 | 0.282 | |
| Shortcut NeuronModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 78.9 | 0.323 | |
| TEDModel=Mistral, Split=fcon, Access=Black-box2026.01 | 77 | 0.521 | |
| DeconIEPModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 66.4 | 0.133 | |
| Short CircuitModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 65.9 | 0.128 | |
| Shortcut NeuronModel=Mistral, Split=fcon, Access=White-box2026.01 | 63.9 | 0.39 | |
| Short CircuitModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 61.2 | 0.146 | |
| baselineModel=Qwen 2.5, Split=fclean, Access=-2026.01 | 53.1 | — | |
| DeconIEPModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 47.6 | 0.009 | |
| baselineModel=LLaMA-3, Split=fclean, Access=-2026.01 | 46.7 | — | |
| DeconIEPModel=Mistral, Split=fcon, Access=White-box2026.01 | 32.2 | 0.073 | |
| Shortcut NeuronModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 29.7 | 0.234 | |
| baselineModel=Mistral, Split=fclean, Access=-2026.01 | 24.9 | — |