Large language models are increasingly used as stand-ins for human users in simulations, but whether they can faithfully reproduce a specific person's behavior and decision logic over time is an open question. A new paper on arXiv investigates this in a longitudinal paper-trading study, according to its abstract.
The paper's title draws a distinction between behavioral accuracy and individual-level fidelity. The authors suggest that standard metrics, which often measure how well a model matches aggregate or average behavior, may not capture whether a simulator actually replicates one person's decision process over time. The study appears to focus on that gap.
Because the abstract is truncated, the results are not described in the available text. The framing, however, points to a need for evaluation methods that assess whether LLM simulators can track an individual's evolving choices, not just produce statistically plausible outputs.