List your company
ResearchFormerly WizWand

Token Attribution Faithfulness on SQuAD v2.0

57.82DisagreementGradient
1.566416.170730.77545.3793Mar 20, 2026
Updated 7mo ago

Evaluation results

MethodLinks
GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.0357.8235.49-22.33
RandomBase Model=Qwen3-4B-Instruct-2507, Method category=Reference2026.0348.89501.11
Input×GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.0346.2347.41.17
RolloutBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.0343.8166.1222.31
IFRBase Model=Qwen3-4B-Instruct-2507, Method category=Contextual mixing-based2026.0343.7466.4622.72
Integrated GradientsBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.0341.3949.998.6
RandomModel=Mistral-7B-Instruct-v0.32026.0334.6334.680.05
IFRModel=Mistral-7B-Instruct-v0.32026.0332.5652.7820.22
GradientModel=Mistral-7B-Instruct-v0.32026.0331.1927.56-3.63
RolloutModel=Mistral-7B-Instruct-v0.32026.0327.9456.8828.94
Last layerBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.0326.0472.4146.37
Input×GradientModel=Mistral-7B-Instruct-v0.32026.0325.6530.444.79
RandomBackbone model=Llama-2-7B-Chat2026.0323.7425.041.3
RolloutBackbone model=Llama-2-7B-Chat2026.0322.6427.354.71
GradientBackbone model=Llama-2-7B-Chat2026.0316.9415.55-1.39
Input×GradientBackbone model=Llama-2-7B-Chat2026.0314.7317.162.43
DePassBase Model=Qwen3-4B-Instruct-2507, Method category=Decomposition-based2026.0313.7466.1952.45
DePassBackbone model=Llama-2-7B-Chat2026.0312.6227.4914.87
IFRBackbone model=Llama-2-7B-Chat2026.0312.6226.292.65
Dual Path Attribution (DPA)Base Model=Qwen3-4B-Instruct-2507, Method category=Proposed Method2026.0311.9393.8681.93
MeanBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.0311.6991.5579.86
MeanModel=Mistral-7B-Instruct-v0.32026.0310.7865.9955.21
Integrated GradientsBackbone model=Llama-2-7B-Chat2026.039.8726.0816.21
Last layerModel=Mistral-7B-Instruct-v0.32026.039.7962.8453.05
Last layerBackbone model=Llama-2-7B-Chat2026.038.931.9523.05
Integrated GradientsModel=Mistral-7B-Instruct-v0.32026.038.6755.9847.31
Dual Path Attribution (DPA)Model=Mistral-7B-Instruct-v0.32026.038.0670.9762.91
DePassModel=Mistral-7B-Instruct-v0.32026.037.9253.9346.01
MeanBackbone model=Llama-2-7B-Chat2026.035.4833.3527.87
Dual Path Attribution (DPA)Backbone model=Llama-2-7B-Chat2026.033.7336.0832.35