Automated Bug Triaging using Instruction-Tuned Large Language Models

2025-08-28

AI Summary

To address low efficiency and poor consistency in bug assignment for large-scale open-source projects, this paper proposes a lightweight LLM-driven framework. The method employs instruction-tuned large language models enhanced with LoRA-based low-rank adaptation and candidate-constrained decoding to ensure outputs are valid developer IDs while enabling human-in-the-loop collaboration. Unlike traditional approaches relying on handcrafted features or graph neural networks, our framework eliminates complex feature engineering, reduces deployment overhead, and achieves efficient inference. Experimental evaluation on the Eclipse JDT and Mozilla datasets demonstrates a Top-10 hit rate of 0.753; notably, accuracy improves significantly on recent project snapshots, confirming the framework’s effectiveness and practical utility in real-world development settings.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Search and Optimization: Learning to Search

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Bug triaging, the task of assigning new issues to developers, is often slow and inconsistent in large projects. We present a lightweight framework that instruction-tuned large language model (LLM) with LoRA adapters and uses candidate-constrained decoding to ensure valid assignments. Tested on EclipseJDT and Mozilla datasets, the model achieves strong shortlist quality (Hit at 10 up to 0.753) despite modest exact Top-1 accuracy. On recent snapshots, accuracy rises sharply, showing the framework's potential for real-world, human-in-the-loop triaging. Our results suggest that instruction-tuned LLMs offer a practical alternative to costly feature engineering and graph-based methods.

Problem

• Automates bug triaging to assign issues efficiently
• Improves slow and inconsistent developer assignment process
• Uses instruction-tuned LLMs to replace costly engineering methods

Innovation

• Instruction-tuned LLMs with LoRA adapters
• Candidate-constrained decoding for valid assignments
• Lightweight framework avoiding costly feature engineering