5th Year Master's Thesis Presentation - Zhijie Xu

— 3:30pm

Location:
In Person - Gates Hillman 9115

Speaker:
ZHIJIE XU , Master's StudentComputer Science DepartmentCarnegie Mellon University

Decomposing Complexity: An LLM-Based Approach to Automating Software Engineering Tasks

Task decomposition in software engineering enables the division of complex engineering tasks into manageable components, facilitating modularization and collaborative development. This thesis investigates the practical application of large language model (LLM) task decomposition in software engineering, focusing on how complex GitHub issues can be systematically broken down into smaller, more solvable subtasks.

To understand existing decomposition practices, this work creates and analyzes a dataset focused on task decomposition for ten Apache projects. The dataset analysis reveals operational patterns of task decomposition in real-world open-source software (OSS) projects, providing insights into how experienced developers naturally decompose complex tasks and identifying key characteristics of effective decomposition strategies.

Building on these insights, a decomposition component is integrated into SWE-agent, enabling structured transformation of GitHub issues into subtasks. This component helps to achieve a 24% performance improvement over the non-decomposed baseline on the SWE-bench verified dataset. The results demonstrate the effectiveness of LLM task decomposition in tackling software engineering problems, potentially improving both automatic system performance and human developer productivity.

Thesis Committee
Carolyn Rosé (Chair)
Michael Hilton

Additional Information 

For More Information:
tracyf@cs.cmu.edu


Add event to Google
Add event to iCal