Abstract: Software bugs do not always appear as obvious crashes. Often, software produces an output that looks reasonable but differs from what developers or users expect. Detecting such cor rectness bugs requires a test oracle: a way to determine whether observed behavior is correct. However, real-world software rarely has a complete specification, and the available specifica tions may themselves be incomplete or incorrect. In this talk, we will focus on how inconsistencies occur between different software artifacts. First, I will present PyZZa, a differential-testing fuzzer for finding behavioral discrepancies be tween CPython and PyPy in the Python standard library. PyZZa automatically generates API call sequences, executes them across both interpreters, and compares their results to uncover cross-interpreter bugs. I will then present DocPrism, a multi-lingual tool that uses large lan guage models to detect inconsistencies between code and method-level documentation. To gether, these projects illustrate two complementary directions: comparing one implementation against another, and comparing an implementation against its documentation. I will conclude with discussing the harder challenge beyond detecting differences—deciding which differences represent real and actionable software problems.

When: Saturday, 8th August, 2026 (2:30 PM)
Where: Graduate Seminar Room (ECE 503), Department of CSE, BUET