counterfactuals

John Schulman 重点介绍了 Adam Karvonen 及其同事的研究，该研究将反事实可模拟性作为提升 AI 解释质量的指标。他们开发了数据集与流水线，训练模型对自身行为生成更优的事后解释，并在留出评估中展现了泛化能力。

This paper derives tighter bounds for multivalued probabilities of causation by incorporating causal information from covariates and mediators, extending prior work from binary settings.