Question Answering on TruthfulQA (Domain-level split, o=3)
92.8AccuracyW/O Decontamination
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| W/O DecontaminationModel=LLaMA-3, Split=f_con2026.01 | 92.8 | 0.476 | |
| TEDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 92.3 | 0.471 | |
| ITDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 89.9 | 0.327 | |
| ITDModel=LLaMA-3, Split=f_con, Access=Black-box2026.01 | 89.9 | 0.447 | |
| W/O DecontaminationModel=Qwen2.5, Split=f_con2026.01 | 88.5 | 0.313 | |
| W/O DecontaminationModel=Mistral, Split=f_con2026.01 | 86.1 | 0.592 | |
| ITDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 86.1 | 0.592 | |
| TEDModel=Qwen2.5, Split=f_con, Access=Black-box2026.01 | 84.6 | 0.274 | |
| Short CircuitModel=Mistral, Split=f_con, Access=White-box2026.01 | 79.3 | 0.524 | |
| TEDModel=Mistral, Split=f_con, Access=Black-box2026.01 | 74 | 0.471 | |
| Shortcut NeuronModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 72.6 | 0.274 | |
| DeconIEPModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 64.1 | 0.069 | |
| Shortcut NeuronModel=Mistral, Split=f_con, Access=White-box2026.01 | 63.6 | 0.367 | |
| Short CircuitModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 62.5 | 0.173 | |
| Short CircuitModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 61.4 | 0.042 | |
| baselineModel=Qwen2.5, Split=f_clean2026.01 | 57.2 | — | |
| DeconIEPModel=Mistral, Split=f_con, Access=White-box2026.01 | 56.5 | 0.296 | |
| DeconIEPModel=LLaMA-3, Split=f_con, Access=White-box2026.01 | 55.3 | 0.101 | |
| baselineModel=LLaMA-3, Split=f_clean2026.01 | 45.2 | — | |
| Shortcut NeuronModel=Qwen2.5, Split=f_con, Access=White-box2026.01 | 39.4 | 0.178 | |
| baselineModel=Mistral, Split=f_clean2026.01 | 26.9 | — |