Original title: Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations
Yi Fei Cheng, Fan Yang, Iremsu Bas, Koichiro Niinuma +2 · September 2026
Computer models powered by large language models are now used to simulate human behavior for policy, training, and testing. However, we do not know if they accurately act like people over long stretches of time. Researchers tested this by creating a way to measure how closely these simulations match real human actions across different timescales and perspectives. They recorded 43 hours of real office activity on video and used that information to guide simulated agents with different setups. They found that feeding the models statistical habits, like common daily routines and transition patterns, made overall action choices look the most realistic. Yet, this approach failed in other ways: it broke daily routines into fragmented pieces and erased personal quirks.
Imagine you're trying to recreate a typical workday based on memory. If you stick strictly to a checklist, you might miss the natural flow and variability of how things actually happen. In this paper, simulators that use pattern-based rules end up with activity patterns close to real ones but break routines into unnatural chunks.
AI simulators of human behavior are used for policy and training, but if they don't match reality, decisions based on them could be flawed. This work provides a way to measure and improve that match.
The case study is limited to a 43-hour dataset of office activity, so findings may not generalize to other domains. The evaluation shows that no single conditioning mechanism captures all aspects of behavioral fidelity.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations — Yi Fei Cheng, Fan Yang, Iremsu Bas et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.