Sentiment Classification on SST2 (Deletion/Insertion Robustness)
0.2943Deletion RobustnessSaliency
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SaliencyBackbone=Llama2026.01 | 0.2943 | 0.1887 | |
| Integrated GradientsBackbone=Llama2026.01 | 0.3064 | 0.1768 | |
| RandomBackbone=Mistral2026.01 | 0.3099 | 0.2287 | |
| RandomBackbone=Llama2026.01 | 0.31 | 0.2237 | |
| SaliencyBackbone=Mistral2026.01 | 0.3124 | 0.1763 | |
| Input×GradientsBackbone=Mistral2026.01 | 0.3145 | 0.1804 | |
| Layer Gradient×ActivationBackbone=Mistral2026.01 | 0.3146 | 0.1806 | |
| DeepliftBackbone=Mistral2026.01 | 0.3153 | 0.1791 | |
| DeepliftBackbone=Llama2026.01 | 0.3279 | 0.1723 | |
| Grad-ELLMBackbone=Llama, Attention Mode=loosen attention (w/)2026.01 | 0.3279 | 0.1891 | |
| Layer Gradient×ActivationBackbone=Llama2026.01 | 0.329 | 0.1722 | |
| Input×GradientsBackbone=Llama2026.01 | 0.3291 | 0.1723 | |
| Integrated GradientsBackbone=Mistral2026.01 | 0.3331 | 0.1783 | |
| Grad-ELLMBackbone=Mistral, Loosen Attention=true2026.01 | 0.3485 | 0.171 | |
| Grad-ELLMBackbone=Mistral, Loosen Attention=false2026.01 | 0.3515 | 0.1744 | |
| AttentionBackbone=Llama2026.01 | 0.3577 | 0.1584 | |
| Grad-ELLMBackbone=Llama, Attention Mode=original attention (w/o)2026.01 | 0.3639 | 0.1687 | |
| Value ZeroingBackbone=Llama2026.01 | 0.3755 | 0.1468 | |
| Value ZeroingBackbone=Mistral2026.01 | 0.3801 | 0.1593 | |
| AttentionBackbone=Mistral2026.01 | 0.3853 | 0.1561 |