Question Answering on TruthfulQA Exact split, o=3
98.6AccuracyW/O Decontamination
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| W/O DecontaminationModel=LLaMA-3, Split=f_con2026.01 | 98.6 | 51.2 | |
| W/O DecontaminationModel=Mistral, Split=f_con2026.01 | 97.6 | 68.9 | |
| ITDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 97.6 | 68.9 | |
| W/O DecontaminationModel=Qwen2.5, Split=f_con2026.01 | 97.6 | 42.1 | |
| ITDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 97.6 | 42.1 | |
| TEDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 96.6 | 49.3 | |
| ITDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 94.3 | 46.9 | |
| Short CircuitModel=Mistral, Split=f_con, Access=White-box2026.01 | 93.3 | 64.6 | |
| TEDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 88 | 32.5 | |
| Shortcut NeuronModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 81.8 | 34.5 | |
| Shortcut NeuronModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 68.4 | 12.9 | |
| Shortcut NeuronModel=Mistral, Split=f_con, Access=White-box2026.01 | 66.8 | 38.1 | |
| Short CircuitModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 65 | 9.5 | |
| DeconIEPModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 62 | 6.5 | |
| Short CircuitModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 59.3 | 12 | |
| baselineModel=Qwen2.5, Split=f_clean2026.01 | 55.5 | — | |
| DeconIEPModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 53.6 | 6.2 | |
| TEDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 53.5 | 24.8 | |
| DeconIEPModel=Mistral, Split=f_con, Access=White-box2026.01 | 53.3 | 24.6 | |
| baselineModel=LLaMA-3, Split=f_clean2026.01 | 47.4 | — | |
| baselineModel=Mistral, Split=f_clean2026.01 | 28.7 | — |