Despite the remarkable performance of neural networks and large language models, these systems remain vulnerable to adversarial attacks and may produce incorrect outputs that are difficult to identify. Existing solutions often require substantial computational resources, introduce high inference overhead, or fail to provide sufficient robustness. To mitigate these shortcomings, our research develops structured algorithms that learn compact programs or informative interactions from data. The first two works use program synthesis to expose and repair vulnerabilities in neural-network classifiers, while the third work extends exact learning to the selection of programs generated by large language models. Together, these works demonstrate that carefully designed program representations and learning procedures can substantially improve the efficiency, robustness, and correctness of modern learning systems.
PhD student under the supervision of Prof. Dana Drachsler Cohen.
