Faithfulness evaluation on BoolQ benchmark leaderboard · SOTA2 Research

Faithfulness evaluation on BoolQ

37AUC π-Soft-NSGrad-ELLM

Evaluation Results

Method	Links
Grad-ELLM2026.01		37	2.462
Random2026.01		33.4	2.381
DeepLIFT2026.01		32.6	1.953
Saliency2026.01		32.4	1.778
Input×Gradients2026.01		31.7	1.765
Layer Gradient×Activation2026.01		31.6	1.832
Grad-ELLM2026.01		31.6	0.451
Grad-ELLM2026.01		31.6	45.1
Random2026.01		29.9	0.487
Random2026.01		29.9	48.7
Saliency2026.01		29.4	0.448
Saliency2026.01		29.4	44.8
Input×Gradients2026.01		29.2	0.452
Input×Gradients2026.01		29.2	45.2
DeepLIFT2026.01		28.9	0.444
DeepLIFT2026.01		28.9	44.4
Layer Gradient×Activation2026.01		28.8	0.448
Layer Gradient×Activation2026.01		28.8	44.8
Integrated Gradients2026.01		26.8	0.422
Integrated Gradients2026.01		26.8	42.2
Integrated Gradients2026.01		26.6	1.434
Attention2026.01		23.6	0.388
Attention2026.01		23.6	38.8
Value Zeroing2026.01		15.8	0.341
Value Zeroing2026.01		15.8	34.1
Attention2026.01		12.1	0.356
Value Zeroing2026.01		11.6	0.33