List your company
ResearchFormerly WizWand

Token Attribution Faithfulness on Known 1000

5.6DistanceDual Path Attribution (DPA)
4.94529.365113.78518.2049Mar 20, 2026
Updated 7mo ago

Evaluation results

MethodLinks
Dual Path Attribution (DPA)Backbone model=Llama-2-7B-Chat2026.035.634.4128.81—
MeanBackbone model=Llama-2-7B-Chat2026.035.9725.119.13—
DePassBase Model=Qwen3-4B-Instruct-2507, Method category=Decomposition-based2026.036.2426.7420.5—
Last layerBackbone model=Llama-2-7B-Chat2026.036.5328.8522.32—
Dual Path Attribution (DPA)Base Model=Qwen3-4B-Instruct-2507, Method category=Proposed Method2026.036.5422.3715.83—
Dual Path Attribution (DPA)Model=Mistral-7B-Instruct-v0.32026.037.1643.2336.07—
DePassBackbone model=Llama-2-7B-Chat2026.037.5526.2918.74—
Integrated GradientsBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.037.5516.188.63—
Last layerModel=Mistral-7B-Instruct-v0.32026.037.830.4122.61—
Last layerBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.038.0322.8214.79—
MeanModel=Mistral-7B-Instruct-v0.32026.038.6132.2723.66—
MeanBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.039.522.3512.85—
DePassModel=Mistral-7B-Instruct-v0.32026.039.6437.5327.89—
Integrated GradientsModel=Mistral-7B-Instruct-v0.32026.039.7238.1428.42—
Dual Path Attribution (DPA)Model=Qwen2.5-32B-Instruct2026.0310.32—31.0341.35
Input×GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.0310.5417.957.41—
Last layerModel=Qwen2.5-32B-Instruct, Category=Attention-based2026.0310.63—21.5532.18
GradientBase Model=Qwen3-4B-Instruct-2507, Method category=Gradient-based2026.0310.8818.077.19—
Input×GradientBackbone model=Llama-2-7B-Chat2026.0311.3120.319—
RandomBase Model=Qwen3-4B-Instruct-2507, Method category=Reference2026.0311.5219.648.12—
Integrated GradientsBackbone model=Llama-2-7B-Chat2026.0311.5825.313.72—
MeanModel=Qwen2.5-32B-Instruct, Category=Attention-based2026.0312.06—21.0133.07
GradientBackbone model=Llama-2-7B-Chat2026.0312.1418.756.61—
DePassModel=Qwen2.5-32B-Instruct, Category=Decomposition-based2026.0312.29—8.2420.53
RandomBackbone model=Llama-2-7B-Chat2026.0313.1720.647.47—
IFRBackbone model=Llama-2-7B-Chat2026.0315.2212.76-2.46—
RolloutBase Model=Qwen3-4B-Instruct-2507, Method category=Attention-based2026.0315.2511.79-3.46—
RolloutBackbone model=Llama-2-7B-Chat2026.0315.4516.170.72—
IFRBase Model=Qwen3-4B-Instruct-2507, Method category=Contextual mixing-based2026.0315.811.28-4.52—
Input×GradientModel=Mistral-7B-Instruct-v0.32026.0316.3128.1811.87—
GradientModel=Mistral-7B-Instruct-v0.32026.0317.3124.667.35—
Integrated GradientsModel=Qwen2.5-32B-Instruct, Category=Gradient-based2026.0318.6—9.127.7
IFRModel=Qwen2.5-32B-Instruct, Category=Contextual mixing-based2026.0318.75—7.8626.61
GradientModel=Qwen2.5-32B-Instruct, Category=Gradient-based2026.0319.44—7.2326.67
RandomModel=Mistral-7B-Instruct-v0.32026.0319.6726.386.71—
RolloutModel=Qwen2.5-32B-Instruct, Category=Attention-based2026.0319.7—-8.3311.37
RolloutModel=Mistral-7B-Instruct-v0.32026.0319.7920.460.67—
RandomModel=Qwen2.5-32B-Instruct2026.0320—9.7929.79
Input × GradientModel=Qwen2.5-32B-Instruct, Category=Gradient-based2026.0321.52—5.426.92
IFRModel=Mistral-7B-Instruct-v0.32026.0321.9716.43-5.54—