Counterfactual Generation on SNLI Premise
0.759LFRFIZLE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| FIZLEExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B2026.01 | 0.759 | 0.908 | 41.54 | |
| Causal What-IfsExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B2026.01 | 0.75 | 0.864 | 41.07 | |
| iFlip-NLExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B, Feedback Type=Natural Language2026.01 | 0.73 | 0.789 | 77 | |
| iFlip-SHAPExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B, Feedback Type=Feature Attribution2026.01 | 0.68 | 0.7 | 57.99 | |
| iFlip-ConfExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B, Feedback Type=Confidence2026.01 | 0.665 | 0.688 | 62.94 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=Confidence2026.01 | 0.66 | 0.854 | 53.72 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=Natural Language2026.01 | 0.638 | 0.879 | 148.41 | |
| CGGExplained Model=RoBERTa-base, Generator Model=LLaMA 3.3-70B2026.01 | 0.635 | 0.893 | 59.23 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=Grad×Input2026.01 | 0.624 | 0.854 | 56.02 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=AttnLRP2026.01 | 0.596 | 0.851 | 55.3 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=Natural Language2026.01 | 0.596 | 0.895 | 68.16 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=SHAP2026.01 | 0.587 | 0.847 | 56.81 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=Confidence2026.01 | 0.58 | 0.763 | 58.65 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=Grad×Input2026.01 | 0.58 | 0.751 | 55.38 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=Natural Language2026.01 | 0.566 | 0.821 | 69.45 | |
| iFlipBackbone LLM=OLMo2-7B, Feedback Type=LIME2026.01 | 0.558 | 0.866 | 58.18 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=LIME2026.01 | 0.558 | 0.759 | 57.87 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=AttnLRP2026.01 | 0.528 | 0.743 | 56.42 | |
| iFlipBackbone LLM=LLaMA3.3-70B, Feedback Type=SHAP2026.01 | 0.522 | 0.743 | 57.62 | |
| BAEExplained Model=RoBERTa-base2026.01 | 0.518 | 0.894 | 103.32 | |
| Causal What-IfsBackbone LLM=Qwen3-32B2026.01 | 0.482 | 0.842 | 36.14 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=Confidence2026.01 | 0.466 | 0.895 | 62.69 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=LIME2026.01 | 0.458 | 0.891 | 62.98 | |
| FIZLEBackbone LLM=Qwen3-32B2026.01 | 0.454 | 0.901 | 38.66 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=SHAP2026.01 | 0.452 | 0.878 | 58.6 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=AttnLRP2026.01 | 0.448 | 0.882 | 58.41 | |
| iFlipBackbone LLM=Qwen3-32B, Feedback Type=Grad×Input2026.01 | 0.44 | 0.881 | 58.66 | |
| Causal What-IfsBackbone LLM=OLMo2-7B2026.01 | 0.413 | 0.854 | 47.27 | |
| FIZLEBackbone LLM=LLaMA3.3-70B2026.01 | 0.402 | 0.903 | 41.54 | |
| Causal What-IfsBackbone LLM=LLaMA3.3-70B2026.01 | 0.372 | 0.816 | 38.97 | |
| CGGBackbone LLM=LLaMA3.3-70B2026.01 | 0.307 | 0.846 | 51.62 | |
| FIZLEBackbone LLM=OLMo2-7B2026.01 | 0.305 | 0.73 | 173.94 | |
| Polyjuice2026.01 | 0.281 | 0.772 | 96.95 | |
| PolyjuiceExplained Model=RoBERTa-base2026.01 | 0.212 | 0.772 | 96.95 | |
| CGGBackbone LLM=Qwen3-32B2026.01 | 0.2 | 0.881 | 50.31 | |
| CGGBackbone LLM=OLMo2-7B2026.01 | 0.169 | 0.882 | 57.03 | |
| BAE2026.01 | 0.135 | 0.966 | 82.98 |