Simulating Users with State Alignment Beats Response Imitation

We build user simulators that accurately reflect real users by generating natural-language latent states aligned with ground-truth responses.

arXiv Read the paper View code Data Models more to come
HumanLM Framework Overview
Fig 1. HumanLM generates responses by first producing latent states (stance, emotion, communication style) that align with ground-truth user behavior, then synthesizing responses from these aligned states.
Abstract

Large Language Models are increasingly used to simulate how specific users respond to any context, enabling more user-centric applications. However, existing user simulators mostly imitate surface-level patterns and language styles, which fails to reflect the underlying state of real users. To address these limitations, we propose HumanLM, a novel training framework which builds user simulators that accurately reflect real users. Our key insight is that we generate natural-language latent states that align with the ground truth responses through reinforcement learning.

Citation
@article{wu2026humanlm,
  title={HUMANLM: Simulating Users with State Alignment Beats Response Imitation},
  url={https://humanlm.stanford.edu/},
  author={Wu, Shirley and Choi, Evelyn and Khatua, Arpandeep and
          Wang, Zhanghan and He-Yueya, Joy and Weerasooriya, Tharindu Cyril and
          Wei, Wei and Yang, Diyi and Leskovec, Jure and Zou, James},
  year={2026}
}