Bug Report Summarization with Large Language Models

Abstract

The unstructured and verbose nature of bug reports often impedes developers from quickly comprehending the context of a problem and fixing underlying issues. While bugreportsummarizationcan facilitate fastercomprehension,existingmethodsoften rely on surface-level textual cues, leading to broken or disorganized summaries and failing to capture deeper semantic nuances. Moreover, these methods often neglect supporting code samples, which are critical for accurately detecting and understand ing software issues. In this work, we propose Chunk-and-Fuse, a novel progressive code integration framework for LLM-based abstractive bug report summarization. Chunk-and-Fuse addresses the challenge of lengthy bug-related code snippets that exceed typical large language model (LLM) context windows by incrementally inte grating segmentedcodeandtextualcontent. Weevaluateourapproachonfourbench mark datasetsacross eightLLMs,achieving7.5%–58.2%improvementsoverextractive baselines and performance comparable to leading abstractive techniques. Our find ings demonstrate that jointly leveraging textual and code information can improve bug comprehension and accelerate software maintenance workflows.