DASA: Fully Gradient-Based Program Analysis (Competition Contribution)

Download PDF
Open Webpage

Felix Mächtle, Jan-Niclas Serr, Nils Loose, Thomas Eisenbarth

Published: 2026, Last Modified: 28 May 2026TACAS (2) 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: We present a novel approach to program analysis that models each instruction using a mathematically differentiable wrapper. This enables gradient-based optimization over program inputs to drive execution towards specific target states, such as assertion violations. Instead of enumerating execution paths or solving symbolic constraints, the program itself is transformed into a differentiable computation graph. By optimizing a loss function that represents the likelihood of reaching a target location, all execution paths are explored simultaneously within a single optimization process.