Visualize: Table1;
Description: ” We propose a gradientguided search over tokens which finds short trigger sequences (e.g., one word for classification and four words for language modeling) that successfully trigger the target prediction. “