Counterfactual-World (CF-World)
收藏资源简介:
Counterfactual-World(CF-World)是由上海交通大学、上海人工智能实验室及香港中文大学联合构建的因果推理评估基准,旨在系统检验文本到图像生成模型在反事实场景下的逻辑演绎能力。该数据集包含1,091个场景组,总计3,273条提示词,涵盖物理、生物、化学、地理和社会学五大核心学科,数据通过人工筛选基础科学原理并结合大语言模型生成。其创新性在于设计了三层渐进式评估框架(事实层、显式反事实层、隐式反事实层),以剥离模型对训练数据统计先验的依赖,专门用于探究模型在违背现实世界规律的假设下进行因果推理与图像合成的潜力。
Counterfactual-World (CF-World) is a causal reasoning evaluation benchmark jointly developed by Shanghai Jiao Tong University, Shanghai AI Laboratory, and The Chinese University of Hong Kong. It aims to systematically evaluate the logical deduction capabilities of text-to-image generation models in counterfactual scenarios. This dataset includes 1,091 scenario groups, totaling 3,273 prompts, covering five core disciplines: physics, biology, chemistry, geography, and sociology. The data is generated by first manually screening basic scientific principles and then leveraging large language models (LLMs). Its core innovation lies in the design of a three-layer progressive evaluation framework, namely the Fact Layer, Explicit Counterfactual Layer, and Implicit Counterfactual Layer, which is intended to eliminate the model's reliance on statistical priors from training data and is specifically dedicated to exploring the potential of models to perform causal reasoning and image synthesis under assumptions that violate the laws of the real world.

- 1Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning上海交通大学; 上海人工智能实验室; 香港中文大学 · 2026年



