This program is tentative and subject to change.
Automated unit test generation has been extensively explored in research, with recent advances highlighting the considerable promise of Large Language Models (LLMs). Models such as GPT-4, trained on extensive corpora of text and code, have demonstrated strong capabilities across various code-related tasks—including unit test generation. Nevertheless, current LLM-based methods tend to operate with a narrow focus, often limited to the immediate code context (e.g., variable references) while overlooking richer, task-specific knowledge sources. For instance, they frequently fail to leverage existing test cases of related methods, which could offer highly relevant guidance. Furthermore, many of these tools emphasize achieving high code coverage, frequently compromising the practical usability, functional correctness, and long-term maintainability of the generated tests.
To address these issues, we introduce a novel mechanism called Reference-Based Retrieval Augmentation, which enhances traditional Retrieval-Augmented Generation (RAG) by incorporating task-aware context retrieval. In the context of unit test generation, we define “test reference relationships” as the potential for test reusability or referential value between a focal method and other methods within the codebase. These relationships allow the system to retrieve pertinent methods and their accompanying unit tests, providing rich contextual clues for generating high-quality tests. Our approach further decomposes test construction into three structured phases—Given, When, and Then—aligning with the typical test design pattern. For each phase, RefTest retrieves and utilizes examples from existing tests of reference methods, offering targeted support for test setup, method invocation, and assertion writing.
We implemented this approach in a tool named RefTest, which systematically conducts preprocessing, test reference retrieval, and unit test generation. An incremental generation strategy is adopted, wherein each newly created test informs and improves subsequent ones. RefTest was evaluated on 12 open-source projects containing 1,515 methods. Results show that it significantly surpasses existing tools across multiple dimensions: correctness, completeness, and maintainability of the generated tests.
This program is tentative and subject to change.
Wed 19 NovDisplayed time zone: Seoul change
14:00 - 15:30 | |||
14:00 10mTalk | LLMs for Automated Unit Test Generation and Assessment in Java: The AgoneTest Framework Research Papers Andrea Lops Polytechnic University of Bari, Italy, Fedelucio Narducci Polytechnic University of Bari, Azzurra Ragone University of Bari, Michelantonio Trizio Wideverse, Claudio Bartolini Wideverse s.r.l. | ||
14:10 10mTalk | µOpTime: Statically Reducing the Execution Time of Microbenchmark Suites Using Stability Metrics Journal-First Track Nils Japke TU Berlin & ECDF, Martin Grambow TU Berlin & ECDF, Christoph Laaber Simula Research Laboratory, David Bermbach TU Berlin | ||
14:20 10mTalk | Reference-Based Retrieval-Augmented Unit Test Generation Journal-First Track Zhe Zhang Beihang University, Liu Xingyu Beihang University, Yuanzhang Lin Beihang University, Xiang Gao Beihang University, Hailong Sun Beihang University, Yuan Yuan Beihang University | ||
14:30 10mTalk | Using Active Learning to Train Predictive Mutation Testing with Minimal Data Research Papers Miklos Borsi Karlsruhe Institute of Technology | ||
14:40 10mTalk | Clarifying Semantics of In-Context Examples for Unit Test Generation Research Papers Chen Yang Tianjin University, Lin Yang Tianjin University, Ziqi Wang Tianjin University, Dong Wang Tianjin University, Jianyi Zhou Huawei Cloud Computing Technologies Co., Ltd., Junjie Chen Tianjin University | ||
14:50 10mTalk | An empirical study of test case prioritization on the Linux Kernel Journal-First Track Haichi Wang College of Intelligence and Computing, Tianjin University, Ruiguo Yu College of Intelligence and Computing, Tianjin University, Dong Wang Tianjin University, Yiheng Du College of Intelligence and Computing, Tianjin University, Yingquan Zhao Tianjin University, Junjie Chen Tianjin University, Zan Wang Tianjin University | ||
15:00 10mTalk | Automated Generation of Issue-Reproducing Tests by Combining LLMs and Search-Based Testing Research Papers Konstantinos Kitsios University of Zurich, Marco Castelluccio Mozilla, Alberto Bacchelli University of Zurich Pre-print | ||
15:10 10mTalk | Using Fourier Analysis and Mutant Clustering to Accelerate DNN Mutation Testing Research Papers | ||
15:20 10mTalk | WEST: Specification-Based Test Generation for WebAssembly Research Papers | ||