Faithfulness Evaluation on IMDb
57.2AUC π-Soft-NSGrad-ELLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Grad-ELLMBackbone=Llama2026.01 | 57.2 | 43 | |
| RandomBackbone=Llama2026.01 | 55.7 | 42.2 | |
| AttentionBackbone=Llama2026.01 | 52.6 | 37.5 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 43.3 | 42.1 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 43.3 | 42.1 | |
| DeepLIFTBackbone=Mistral2026.01 | 43.2 | 41.5 | |
| DeepLIFTBackbone=Mistral2026.01 | 43.2 | 41.5 | |
| Input×GradientsBackbone=Mistral2026.01 | 43.1 | 42.2 | |
| Input×GradientsBackbone=Mistral2026.01 | 43.1 | 42.2 | |
| AttentionBackbone=Mistral2026.01 | 42.5 | 34.6 | |
| AttentionBackbone=Mistral2026.01 | 42.5 | 34.6 | |
| SaliencyBackbone=Llama2026.01 | 42 | 29.6 | |
| Integrated GradientsBackbone=Llama2026.01 | 41.2 | 29.3 | |
| SaliencyBackbone=Mistral2026.01 | 40.2 | 38.8 | |
| SaliencyBackbone=Mistral2026.01 | 40.2 | 38.8 | |
| DeepLIFTBackbone=Llama2026.01 | 40.1 | 28.1 | |
| Grad-ELLMBackbone=Mistral2026.01 | 40 | 53 | |
| Grad-ELLMBackbone=Mistral2026.01 | 40 | 53 | |
| Value ZeroingBackbone=Llama2026.01 | 39.8 | 27.1 | |
| Input×GradientsBackbone=Llama2026.01 | 38.4 | 27.3 | |
| Layer Gradient×ActivationBackbone=Llama2026.01 | 38.4 | 27.2 | |
| Value ZeroingBackbone=Mistral2026.01 | 36.9 | 6.5 | |
| Value ZeroingBackbone=Mistral2026.01 | 36.9 | 6.5 | |
| Integrated GradientsBackbone=Mistral2026.01 | 36.2 | 35.7 | |
| Integrated GradientsBackbone=Mistral2026.01 | 36.2 | 35.7 | |
| RandomBackbone=Mistral2026.01 | 29.3 | 61.8 | |
| RandomBackbone=Mistral2026.01 | 29.3 | 61.8 |