Saliency maps derived from input gradients are popular for visualizing and controlling model responses [1,
2, 3]. In RF neuromodulation or analogous control settings, however, raw gradients often exhibit speckled,
high-frequency artifacts that are hard to actuate. We propose a structured-gradient formulation that imposes
spatial coherence and sparsity on the raw gradient while preserving fidelity to the model’s objective under
stimulation [4, 5]. We show (Fig. ??, ??) that norm-regularized gradients reduce speckle and improve region
targeting according to standard perturbation tests.