Faithfulness Evaluation on ERASER IOI prompts, n=50
100SufficiencyOurs (Circuit)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Ours (Circuit)Explanation Type=Circuit-based2026.02 | 100 | 22 | 36 | |
| RandomExplanation Type=Random heads2026.02 | 46.7 | 28.8 | 35.6 | |
| Attention-basedExplanation Type=Attention weights2026.02 | 16.7 | 26.6 | 20.6 |