TRUST-REGION SALIENCY-GUIDED LOCAL SEARCH FOR INTERPRETABLE SEQUENCE DESIGN AT FIXED EDIT BUDGETS

Discrete sequence design under a fixed edit budget can match target model outputs, but often returns dispersed, multi-cluster edits that are hard to interpret. We present $\textbf{SAGE-TRSwap}$, a saliency-guided trust-region local search that optimizes the same prediction loss as a Ledidi-style relaxation+pruning baseline while biasing proposals toward high-attribution regions and enabling budget-preserving swap refinements. Across 12 regulatory targets/tracks and 5 random starts per target (60 runs per budget), SAGE-TRSwap sharply reduces edit span and cluster count at all budgets while maintaining or improving absolute error.