Can LLMS reproduce the treatment effects of field experiments in silico? Interesting new paper suggests they do so about 78% of the time: arxiv.org/pdf/2504.01167