Sentiment Classification on IMDb (Robustness Metrics)
17.36Deletion RateSaliency
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SaliencyBackbone=Llama2026.01 | 17.36 | 13.16 | |
| Integrated GradientsBackbone=Llama2026.01 | 17.69 | 10.98 | |
| Integrated GradientsBackbone=Mistral2026.01 | 18.29 | 10.11 | |
| RandomBackbone=Llama2026.01 | 18.44 | 16.96 | |
| RandomBackbone=Mistral2026.01 | 18.75 | 16.14 | |
| DeepliftBackbone=Llama2026.01 | 20.21 | 11.89 | |
| SaliencyBackbone=Mistral2026.01 | 20.24 | 13.16 | |
| Input×GradientsBackbone=Llama2026.01 | 20.31 | 12.2 | |
| Layer Gradient×ActivationBackbone=Llama2026.01 | 20.31 | 12.21 | |
| DeepliftBackbone=Mistral2026.01 | 20.56 | 8.88 | |
| Input×GradientsBackbone=Mistral2026.01 | 20.59 | 8.95 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 20.6 | 8.97 | |
| Grad-ELLMBackbone=Llama, Attention Mode=loosen attention (w/)2026.01 | 22.37 | 12.04 | |
| Grad-ELLMBackbone=Mistral, Loosen Attention=false2026.01 | 24.07 | 12.01 | |
| Grad-ELLMBackbone=Mistral, Loosen Attention=true2026.01 | 24.52 | 9.78 | |
| Value ZeroingBackbone=Mistral2026.01 | 25.1 | 8.52 | |
| Grad-ELLMBackbone=Llama, Attention Mode=original attention (w/o)2026.01 | 26.6 | 9.3 | |
| AttentionBackbone=Llama2026.01 | 26.91 | 9.17 | |
| Value ZeroingBackbone=Llama2026.01 | 27.29 | 8.51 | |
| AttentionBackbone=Mistral2026.01 | 28.53 | 6.36 |