Google demonstrated that Gradient L2 salience leads to consistently better token salience explanations than other explainability models in LIT visualizations (the Language Interpretability Tool). https://ai.googleblog.com/2022/12/will-you-find-these-shortcuts.html