Semester of Graduation
Summer 2026
Degree Type
Thesis
Degree Name
Master of Science in Artificial Intelligence
Department
Computer Science - College of Computing and Software Engineering
Committee Chair/First Advisor
Md. Abdullah Al Hafiz Khan
Second Advisor
Yong Pei
Third Advisor
Chih-Cheng Hung
Abstract
Large language models are least available where they could do the most good. Hospitals, clinics, and schools operate on aging or outdated local hardware, and privacy law make remote APIs a non-option, meaning compression is the only viable option left. Most compression methods fail to preserve the domain-specific knowledge needed by these settings, and those that find sparse task-specific subnetworks discard any learned structure, producing only a compressed model. We ask whether that structure can instead be extracted once, securely, and reused. We propose PRIZE (Performance-Retaining Importance-based transferable Zero-mask Extraction), a structured-pruning framework whose deliverable is a reusable binary mask, a few kilobytes marking which structural units of each transformer sublayer a task needs, carrying no weights and no data. It is learned by instrumenting a pretrained model with sigmoid gates and jointly optimizing task loss and sparsity, while backbone gradients pass through null-space projection, clipping, and calibrated noise. Across four model families, masks transfer to fresh checkpoints and recover near-dense accuracy in one short fine-tune, and leakage falls on three of four models. An audit establishes three qualifications we report rather than minimize: the gates act on each sublayer's output rather than on attention heads and feed-forward channels, the trained gates settle near one half rather than separating, so the export threshold sets the mask's sparsity, and the learned-versus-random comparison, re-indexed by attained sparsity, confirms no advantage. Excision is nonetheless implemented and verified equivalent, giving reductions of 37-68% in parameters. Transfer, privacy, and excision results are unaffected.