Program Comprehension for Human and LLM-Based Software Maintenance

Loading...
Thumbnail Image

TR Number

Date

2026-10-06

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Modern software systems increasingly rely on abstractions, transformations, and automated tools that make it difficult to understand how source-level programs relate to their actual behavior. Effective debugging and testing therefore depend on accurate program comprehension by both human and automated agents that perform these tasks. This thesis investigates how program comprehension can be improved for developers and systematically assessed for Large Language Models (LLMs) used in software maintenance. The first part focuses on improving human program comprehension in distributed systems. This thesis develops an interactive debugging approach that connects developer-written SQL to its transformed distributed execution by mapping intermediate execution data back to corresponding parts of the original query. This allows developers to understand complex execution behavior using the same abstractions in which the program was originally expressed. As LLMs are increasingly used for coding tasks, program comprehension shifts from human developers to automated agents. The second part investigates LLM program comprehension in automated debugging. The evaluation examines fault localization under semantics-preserving code changes to determine whether LLMs consistently understand program behavior or are influenced by superficial properties of how code is written. The final part extends this investigation to LLM-based test generation under software evolution, evaluating whether generated tests remain reliable when program functionality is preserved and whether they appropriately reflect changes when program behavior is modified. Together, these contributions examine program comprehension from two complementary perspectives: improving developers' understanding of program behavior and evaluating whether LLMs possess the comprehension required for reliable automated debugging and testing.

Description

Keywords

Debugging, Data-Intensive Scalable Computing, Large Language Models

Citation