PRMBENCH
收藏资源简介:
PRMBENCH是由复旦大学、苏州大学、上海人工智能实验室等机构联合开发的一个细粒度基准测试数据集,旨在评估过程级奖励模型(PRMs)在复杂推理任务中的表现。该数据集包含6216个问题实例和83456个步骤级标签,涵盖了简洁性、合理性和敏感性等多个评估维度。数据集的创建过程结合了人工和自动生成的方法,确保了数据的多样性和质量。PRMBENCH的应用领域主要集中在复杂推理任务的模型评估,旨在解决现有基准测试在评估PRMs时无法全面覆盖错误类型和场景的问题,为未来研究提供了重要的参考和指导。
PRMBENCH is a fine-grained benchmark dataset jointly developed by institutions including Fudan University, Soochow University, Shanghai AI Laboratory and others. It aims to evaluate the performance of process-level reward models (PRMs) on complex reasoning tasks. This dataset contains 6216 question instances and 83456 step-level labels, covering multiple evaluation dimensions such as conciseness, rationality and sensitivity. The dataset was created using a combination of manual and automated generation methods to ensure the diversity and quality of the data. The application fields of PRMBENCH mainly focus on model evaluation for complex reasoning tasks, aiming to address the issue that existing benchmark tests fail to comprehensively cover error types and scenarios when evaluating PRMs, thus providing important references and guidance for future research.

- 1PRMBench: A Fine-grained and Challenging Benchmark for Process-Level Reward Models复旦大学, 苏州大学, 上海人工智能实验室, 石溪大学, 香港中文大学 · 2025年



