Large language models are increasingly cast as synthetic survey respondents, but their usefulness depends on whether they can reproduce the actual choices of the people they stand in for. A new arXiv preprint asks which input works better: a description of who a person is, or a record of what that person has done.
The paper, titled "Behavioral History Outperforms Descriptions of the Person for LLM Synthetic Personas," directly addresses this comparison. According to the abstract, the validity of such personas rests on whether they reproduce individuals' decisions, and the reported finding is that behavioral history outperforms descriptions of the person.
Because the source abstract is truncated, the article does not include details on the experimental setup, dataset, or effect sizes. The preprint is available on arXiv under identifier 2610.03998.