Faithfulness evaluation on WikiBio
2.3AUC π-Soft-NSHETA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| HETAModel=Phi-3-Medium-14B2026.04 | 2.3 | 4.2 | |
| HETABackbone=Qwen2.5 3B2026.04 | 2.2 | 3.9 | |
| HETAModel=GPT-J 6B2026.04 | 2.2 | 3.8 | |
| HETAModel=LLaMA-3.1 70B2026.04 | 2.1 | 3.7 | |
| Peering (PML)Model=LLaMA-3.1 70B2026.04 | 1.2 | 0.15 | |
| Integrated GradientsModel=LLaMA-3.1 70B2026.04 | 1.15 | 0.13 | |
| Peering (PML)Backbone=Qwen2.5 3B2026.04 | 0.85 | 1.6 | |
| Peering (PML)Model=GPT-J 6B2026.04 | 0.83 | 1.5 | |
| ContextCiteBackbone=Qwen2.5 3B2026.04 | 0.79 | 1.45 | |
| Integrated GradientsModel=GPT-J 6B2026.04 | 0.77 | 1.38 | |
| Peering (PML)Model=Phi-3-Medium-14B2026.04 | 0.76 | 1.48 | |
| Integrated GradientsModel=Phi-3-Medium-14B2026.04 | 0.7 | 1.35 | |
| Attention RolloutModel=LLaMA-3.1 70B2026.04 | 0.61 | -1.48 | |
| TDD-backwardModel=LLaMA-3.1 70B2026.04 | 0.59 | -0.02 | |
| ContextCiteModel=LLaMA-3.1 70B2026.04 | 0.57 | 0.85 | |
| TDD-backwardModel=GPT-J 6B2026.04 | 0.51 | 0.11 | |
| TDD-backwardModel=Phi-3-Medium-14B2026.04 | 0.5 | 0.1 | |
| TDD-backwardBackbone=Qwen2.5 3B2026.04 | 0.49 | 0.12 | |
| Attention RolloutModel=GPT-J 6B2026.04 | 0.46 | 1.91 | |
| Grad-ELLMBackbone=Mistral2026.01 | 0.438 | 0.477 | |
| Grad-ELLMBackbone=Mistral2026.01 | 0.438 | 0.477 | |
| Attention RolloutBackbone=Qwen2.5 3B2026.04 | 0.43 | 1.85 | |
| Attention RolloutModel=Phi-3-Medium-14B2026.04 | 0.43 | 1.85 | |
| RandomBackbone=Mistral2026.01 | 0.402 | 0.512 | |
| RandomBackbone=Mistral2026.01 | 0.402 | 0.512 | |
| ReAgentBackbone=Qwen2.5 3B2026.04 | 0.4 | 1.25 | |
| ReAgentModel=Phi-3-Medium-14B2026.04 | 0.4 | 1.25 | |
| ReAgentModel=GPT-J 6B2026.04 | 0.39 | 1.22 | |
| ReAgentModel=LLaMA-3.1 70B2026.04 | 0.38 | 1.15 | |
| SEA-CoTModel=GPT-J 6B2026.04 | 0.35 | 1.1 | |
| DeepLIFTBackbone=Mistral2026.01 | 0.348 | 0.391 | |
| DeepLIFTBackbone=Mistral2026.01 | 0.348 | 0.391 | |
| Input×GradientsBackbone=Mistral2026.01 | 0.345 | 0.392 | |
| Input×GradientsBackbone=Mistral2026.01 | 0.345 | 0.392 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 0.344 | 0.389 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 0.344 | 0.389 | |
| SEA-CoTBackbone=Qwen2.5 3B2026.04 | 0.34 | 1.15 | |
| SEA-CoTModel=Phi-3-Medium-14B2026.04 | 0.34 | 1.15 | |
| SaliencyBackbone=Mistral2026.01 | 0.326 | 0.376 | |
| SaliencyBackbone=Mistral2026.01 | 0.326 | 0.376 | |
| AttentionBackbone=Mistral2026.01 | 0.325 | 0.41 | |
| AttentionBackbone=Mistral2026.01 | 0.325 | 0.41 | |
| SEA-CoTModel=LLaMA-3.1 70B2026.04 | 0.31 | 1.05 | |
| Integrated GradientsBackbone=Mistral2026.01 | 0.304 | 0.334 | |
| Integrated GradientsBackbone=Mistral2026.01 | 0.304 | 0.334 | |
| Progressive InferenceModel=LLaMA-3.1 70B2026.04 | 0.26 | 0.95 | |
| Progressive InferenceModel=GPT-J 6B2026.04 | 0.22 | 0.99 | |
| Value ZeroingBackbone=Mistral2026.01 | 0.218 | 0.351 | |
| Value ZeroingBackbone=Mistral2026.01 | 0.218 | 0.351 | |
| Progressive InferenceBackbone=Qwen2.5 3B2026.04 | 0.21 | 1 | |
| Progressive InferenceModel=Phi-3-Medium-14B2026.04 | 0.21 | 1 | |
| Grad-ELLMBackbone=Llama2026.01 | 0.194 | 1.059 | |
| Integrated GradientsBackbone=Qwen2.5 3B2026.04 | 0.18 | 1.1 | |
| RandomBackbone=Llama2026.01 | 0.175 | 1.475 | |
| Integrated GradientsBackbone=Llama2026.01 | 0.165 | 1.003 | |
| DeepLIFTBackbone=Llama2026.01 | 0.162 | 1.007 | |
| AttentionBackbone=Llama2026.01 | 0.158 | 0.571 | |
| Input×GradientsBackbone=Llama2026.01 | 0.149 | 1.829 | |
| Layer Gradient×ActivationBackbone=Llama2026.01 | 0.148 | 0.845 | |
| SaliencyBackbone=Llama2026.01 | 0.139 | 0.66 | |
| Value ZeroingBackbone=Llama2026.01 | 0.106 | 2.396 | |
| fAMLModel=GPT-J 6B2026.04 | -0.02 | 0.21 | |
| fAMLBackbone=Qwen2.5 3B2026.04 | -0.04 | 0.2 | |
| fAMLModel=Phi-3-Medium-14B2026.04 | -0.04 | 0.2 | |
| ContextCiteModel=Phi-3-Medium-14B2026.04 | -0.06 | 0.52 | |
| fAMLModel=LLaMA-3.1 70B2026.04 | -0.07 | 0.46 | |
| ContextCiteModel=GPT-J 6B2026.04 | -0.08 | 0.49 |