Question Answering on TruthfulQA o=1 (Domain-level split)
88.5AccuracyW/O Decontamination
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| W/O DecontaminationModel=LLaMA-3, Split=fcon, Access=-2026.01 | 88.5 | 0.433 | |
| W/O DecontaminationModel=Mistral, Split=fcon, Access=-2026.01 | 86.5 | 0.596 | |
| ITDModel=Mistral, Split=fcon, Access=Black-box2026.01 | 86.5 | 0.596 | |
| ITDModel=LLaMA-3, Split=fcon, Access=Black-box2026.01 | 85.6 | 0.404 | |
| TEDModel=LLaMA-3, Split=fcon, Access=Black-box2026.01 | 85.1 | 0.399 | |
| Short CircuitModel=Mistral, Split=fcon, Access=White-box2026.01 | 82.7 | 0.558 | |
| W/O DecontaminationModel=Qwen 2.5, Split=fcon, Access=-2026.01 | 78.4 | 0.212 | |
| ITDModel=Qwen 2.5, Split=fcon, Access=Black-box2026.01 | 77.4 | 0.202 | |
| TEDModel=Qwen 2.5, Split=fcon, Access=Black-box2026.01 | 76 | 0.188 | |
| TEDModel=Mistral, Split=fcon, Access=Black-box2026.01 | 73.6 | 0.467 | |
| DeconIEPModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 65.5 | 0.083 | |
| Shortcut NeuronModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 64.4 | 0.192 | |
| Shortcut NeuronModel=Mistral, Split=fcon, Access=White-box2026.01 | 64 | 0.371 | |
| Short CircuitModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 63.5 | 0.183 | |
| Short CircuitModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 60.9 | 0.037 | |
| baselineModel=Qwen 2.5, Split=fclean, Access=-2026.01 | 57.2 | — | |
| DeconIEPModel=LLaMA-3, Split=fcon, Access=White-box2026.01 | 46.1 | 0.009 | |
| baselineModel=LLaMA-3, Split=fclean, Access=-2026.01 | 45.2 | — | |
| DeconIEPModel=Mistral, Split=fcon, Access=White-box2026.01 | 40.8 | 0.139 | |
| Shortcut NeuronModel=Qwen 2.5, Split=fcon, Access=White-box2026.01 | 27.4 | 0.298 | |
| baselineModel=Mistral, Split=fclean, Access=-2026.01 | 26.9 | — |