Token Attribution Faithfulness on IMDb
46.75DistanceMean
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MeanBackbone model=Llama-2-7B-Chat2026.03 | 46.75 | 78.6 | 31.85 | |
| Dual Path Attribution (DPA)Backbone model=Llama-2-7B-Chat2026.03 | 48.21 | 74.53 | 26.32 | |
| Integrated GradientsBackbone model=Llama-2-7B-Chat2026.03 | 52.3 | 68.19 | 15.89 | |
| GradientBackbone model=Llama-2-7B-Chat2026.03 | 55.16 | 63.9 | 8.74 | |
| RandomBackbone model=Llama-2-7B-Chat2026.03 | 58.98 | 57.8 | -1.18 | |
| Last layerBackbone model=Llama-2-7B-Chat2026.03 | 59.81 | 60.94 | 1.13 | |
| DePassBackbone model=Llama-2-7B-Chat2026.03 | 68.19 | 56.13 | -12.06 | |
| Input×GradientBackbone model=Llama-2-7B-Chat2026.03 | 69.39 | 50.85 | -18.54 | |
| RolloutBackbone model=Llama-2-7B-Chat2026.03 | 83.31 | 55.54 | -27.77 | |
| IFRBackbone model=Llama-2-7B-Chat2026.03 | 91.62 | 48.67 | -42.95 |