Question Answering on TruthfulQA Semantic-level split o=3
98.1AccuracyW/O Decontamination
Evaluation results
| Method | Links | ||
|---|---|---|---|
| W/O DecontaminationModel=LLaMA-3, Split=f_con2026.01 | 98.1 | 0.514 | |
| W/O DecontaminationModel=Mistral, Split=f_con2026.01 | 97.1 | 0.722 | |
| ITDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 97.1 | 0.722 | |
| ITDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 95.7 | 0.49 | |
| W/O DecontaminationModel=Qwen2.5, Split=f_con2026.01 | 95.2 | 0.421 | |
| TEDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 95.1 | 0.485 | |
| ITDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 94.7 | 0.416 | |
| Short CircuitModel=Mistral, Split=f_con, Access=White-box2026.01 | 91.4 | 0.665 | |
| TEDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 90 | 0.369 | |
| Shortcut NeuronModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 85.2 | 0.385 | |
| Shortcut NeuronModel=Mistral, Split=f_con, Access=White-box2026.01 | 63.4 | 0.385 | |
| DeconIEPModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 63.2 | 0.101 | |
| Short CircuitModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 62.2 | 0.091 | |
| Short CircuitModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 60.8 | 0.141 | |
| TEDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 57.9 | 0.33 | |
| DeconIEPModel=Mistral, Split=f_con, Access=White-box2026.01 | 55.1 | 0.302 | |
| baselineModel=Qwen2.5, Split=f_clean2026.01 | 53.1 | — | |
| Shortcut NeuronModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 52.6 | 0.005 | |
| DeconIEPModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 51.9 | 0.053 | |
| baselineModel=LLaMA-3, Split=f_clean2026.01 | 46.7 | — | |
| baselineModel=Mistral, Split=f_clean2026.01 | 24.9 | — |