Setting

For RQ2,we conduct an ablation study to assess the contri- butions of the word-level sensitivity analysis and sensitivity-aware mutation.First,we implement a random strategy (Random)that randomly selects a word for mutation,i.e., without sensitivity-aware mutation.The candidate substitution list for it is the union of the substitution lists for dirty words and discrepant words (DirLis and DisLis).To demonstrate the importance of sensitivity-aware mutation,we configure two variants of TokenProber:one without Dirtiness-preserving mutation (w/o Dirty) and another without Discrepancy-away mutation (w/o Discrepancy). Additionally,we introduce two specialized versions for a more granular examination of our mutation approach:TokenProber with Dirtiness-Away Mutation,which selects substitutions that differ from the original dirty words (Dirt-away), and TokenProber with Discrepancy- preserving mutation,which selects substitutions similar to the original discrepant words (Discr-prev).Dirt-away and Discr-prev mutate the dirty words and discrepant words in the opposite direction to TokenProber,respectively.

Results

Both dirtiness-preserving mutation and discrepancy-away mutation play significant roles in enhancing the performance of TokenProber