arXiv: Learning user simulators with Turing rewards to train and evaluate agent assistants

arXiv·medium signal

A new arXiv paper (2606.19336) proposes learning human-user simulators using 'Turing rewards' to make simulated users more realistic, aimed at training agent assistants, evaluating personalization systems, and stress-testing interactive agents. Realistic, scalable user simulation is a persistent bottleneck for agent RL and eval — most teams fall back on brittle scripted personas. If the Turing-reward approach holds up, it offers a cheaper synthetic-interaction loop for builders who can't run large human-in-the-loop studies.