Factual Knowledge on Known 1000
5.26Disagreement RateLast layer
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Last layerModel=Llama-3.1-8B-Instruct2026.03 | 5.26 | 24.83 | 19.57 | |
| Dual Path Attribution (DPA)Model=Llama-3.1-8B-Instruct2026.03 | 5.79 | 33.88 | 28.09 | |
| MeanModel=Llama-3.1-8B-Instruct2026.03 | 6.57 | 25.78 | 19.21 | |
| DePassModel=Llama-3.1-8B-Instruct2026.03 | 9.17 | 25.45 | 16.28 | |
| Integrated GradientsModel=Llama-3.1-8B-Instruct2026.03 | 9.92 | 24.49 | 14.57 | |
| GradientModel=Llama-3.1-8B-Instruct2026.03 | 12.14 | 22.94 | 10.8 | |
| Input X GradientModel=Llama-3.1-8B-Instruct2026.03 | 12.58 | 21.27 | 8.69 | |
| RolloutModel=Llama-3.1-8B-Instruct2026.03 | 13.26 | 15.72 | 2.46 | |
| RandomModel=Llama-3.1-8B-Instruct2026.03 | 14.22 | 22.78 | 8.56 | |
| IFRModel=Llama-3.1-8B-Instruct2026.03 | 16.15 | 11.88 | -4.27 |