一句话:提出自动生成测试套件以评估大语言模型潜在偏见的方法,测试由提示和检测偏见的 oracle 组成。
大语言模型正被快速集成到各类软件应用中,可能引入性别、年龄、族裔等偏见,影响应用行为。为应对该挑战,本文探索自动生成测试套件以评估 LLM 的潜在偏见。方法将每个测试定义为两部分:作为输入的提示(prompt)和用于分析 LLM 输出的测试 oracle;oracle 判断输出中是否存在偏见。通过自动生成提示与相应 oracle,目标是系统化、可扩展地开展 LLM 偏见测试。摘要未报告具体实验结果,主要贡献是提出自动化偏见测试的框架与方向。
Large Language Models (LLMs) are being quickly integrated in a myriad of software applications. This may introduce a number of biases, such as gender, age or ethnicity, in the behavior of such applications. To face this challenge, we explore the automatic generation of tests suites to assess the potential biases of an LLM. Each test is defined as a prompt used as input to the LLM and a test oracle that analyses the LLM output to detect the presence of biases.