InterveneBench
收藏资源简介:
InterveneBench是由复旦大学和上海创新研究院联合开发的基准数据集,旨在评估大型语言模型在现实社会政策干预下的因果推理能力。该数据集包含744项来自高质量社会科学研究的实例,覆盖多样化的政策领域,每条实例均源自实证研究,要求模型在不依赖预定义因果图或结构方程的情况下进行推理。数据集通过多智能体流程构建,经过专家验证确保研究级质量。其核心应用是推动开放式的社会科学因果推断,解决现有模型在真实政策场景中结构不可知和干预中心化推理的局限性。
InterveneBench is a benchmark dataset jointly developed by Fudan University and the Shanghai Institute of Innovation, designed to evaluate the causal reasoning abilities of large language models under real-world social policy interventions. This dataset contains 744 instances sourced from high-quality social science research, spanning diverse policy domains. Each instance originates from empirical studies, requiring models to perform causal reasoning without relying on pre-defined causal graphs or structural equations. The dataset is constructed through a multi-agent workflow and validated by experts to ensure research-grade quality. Its core application is to facilitate open-ended causal inference in social science, addressing the limitations of existing models in real policy scenarios: unknown causal structures and intervention-centric reasoning.




