FaithformBench
收藏资源简介:
FaithformBench是一个专为评估数学链式思维自动形式化系统忠实度而设计的基准数据集,由南洋理工大学、牛津大学和爱丁堡大学的研究团队联合创建。该数据集包含12,784个推理步骤及其对应的扰动版本,涵盖四个数学数据集,难度逐级递增,每个步骤均被标注为有效或无效。数据集的构建基于ProcessBench中的有效推理步骤,通过自动扰动技术生成无效步骤,并使用证明助手验证其有效性。该基准旨在系统性地检测自动形式化系统的错误诱导与静默修正两种失败模式,从而评估系统是否忠实保留了输入的逻辑有效性或无效性,为形式化验证领域的可靠性研究提供关键支撑。
FaithformBench is a benchmark dataset specifically designed for evaluating the faithfulness of automatic mathematical chain-of-thought formalization systems, jointly created by research teams from Nanyang Technological University, University of Oxford, and University of Edinburgh. This dataset contains 12,784 reasoning steps and their corresponding perturbed variants, covering four mathematical datasets with gradually increasing difficulty, where each step is labeled as either valid or invalid. The dataset is constructed based on valid reasoning steps sourced from ProcessBench, with invalid steps generated via automatic perturbation techniques and their validity verified using proof assistants. This benchmark aims to systematically detect two failure modes of automatic formalization systems: error induction and silent correction, thereby assessing whether the system faithfully preserves the logical validity or invalidity of the input, providing critical support for reliability research in the field of formal verification.
FaithformBench 数据集概述
FaithformBench 是一个用于评测数学链式思维(Chain-of-Thought)自动形式化忠实度的基准数据集,相关代码已在 GitHub 上开源。该数据集对应的论文为《FaithformBench: Benchmarking Faithfulness of Mathematical Chain-of-Thought Autoformalisation》,目前论文及数据集相关资源将陆续发布。
主要内容
- 研究领域:数学自然语言推理的自动形式化,特别是链式思维过程的忠实度评估。
- 核心任务:通过基准测试,衡量模型在进行数学问题链式思维自动形式化时,输出是否忠实于原始推理过程。
- 当前状态:代码已公开,论文及数据集完整版本标注为“即将发布”(To be released soon)。
资源与访问
- 数据地址:https://github.com/Ighina/FaithformBench
- 内容类型:基准测试数据集 + 相关代码实现





