Token Attribution Faithfulness on SQuAD v2.0
57.82DisagreementGradient
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.03 | 57.82 | 35.49 | -22.33 | |
| RandomBase Model=Qwen3-4B-Instruct-2507, Method category=Reference2026.03 | 48.89 | 50 | 1.11 | |
| Input×GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.03 | 46.23 | 47.4 | 1.17 | |
| RolloutBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.03 | 43.81 | 66.12 | 22.31 | |
| IFRBase Model=Qwen3-4B-Instruct-2507, Method category=Contextual mixing-based2026.03 | 43.74 | 66.46 | 22.72 | |
| Integrated GradientsBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.03 | 41.39 | 49.99 | 8.6 | |
| RandomModel=Mistral-7B-Instruct-v0.32026.03 | 34.63 | 34.68 | 0.05 | |
| IFRModel=Mistral-7B-Instruct-v0.32026.03 | 32.56 | 52.78 | 20.22 | |
| GradientModel=Mistral-7B-Instruct-v0.32026.03 | 31.19 | 27.56 | -3.63 | |
| RolloutModel=Mistral-7B-Instruct-v0.32026.03 | 27.94 | 56.88 | 28.94 | |
| Last layerBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.03 | 26.04 | 72.41 | 46.37 | |
| Input×GradientModel=Mistral-7B-Instruct-v0.32026.03 | 25.65 | 30.44 | 4.79 | |
| RandomBackbone model=Llama-2-7B-Chat2026.03 | 23.74 | 25.04 | 1.3 | |
| RolloutBackbone model=Llama-2-7B-Chat2026.03 | 22.64 | 27.35 | 4.71 | |
| GradientBackbone model=Llama-2-7B-Chat2026.03 | 16.94 | 15.55 | -1.39 | |
| Input×GradientBackbone model=Llama-2-7B-Chat2026.03 | 14.73 | 17.16 | 2.43 | |
| DePassBase Model=Qwen3-4B-Instruct-2507, Method category=Decomposition-based2026.03 | 13.74 | 66.19 | 52.45 | |
| DePassBackbone model=Llama-2-7B-Chat2026.03 | 12.62 | 27.49 | 14.87 | |
| IFRBackbone model=Llama-2-7B-Chat2026.03 | 12.62 | 26.29 | 2.65 | |
| Dual Path Attribution (DPA)Base Model=Qwen3-4B-Instruct-2507, Method category=Proposed Method2026.03 | 11.93 | 93.86 | 81.93 | |
| MeanBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.03 | 11.69 | 91.55 | 79.86 | |
| MeanModel=Mistral-7B-Instruct-v0.32026.03 | 10.78 | 65.99 | 55.21 | |
| Integrated GradientsBackbone model=Llama-2-7B-Chat2026.03 | 9.87 | 26.08 | 16.21 | |
| Last layerModel=Mistral-7B-Instruct-v0.32026.03 | 9.79 | 62.84 | 53.05 | |
| Last layerBackbone model=Llama-2-7B-Chat2026.03 | 8.9 | 31.95 | 23.05 | |
| Integrated GradientsModel=Mistral-7B-Instruct-v0.32026.03 | 8.67 | 55.98 | 47.31 | |
| Dual Path Attribution (DPA)Model=Mistral-7B-Instruct-v0.32026.03 | 8.06 | 70.97 | 62.91 | |
| DePassModel=Mistral-7B-Instruct-v0.32026.03 | 7.92 | 53.93 | 46.01 | |
| MeanBackbone model=Llama-2-7B-Chat2026.03 | 5.48 | 33.35 | 27.87 | |
| Dual Path Attribution (DPA)Backbone model=Llama-2-7B-Chat2026.03 | 3.73 | 36.08 | 32.35 |