MIMESIS: a 9B user simulator as a training environment for agents
The paper introduces MIMESIS, a purpose-built 9B user simulator trained on human conversations with explicit reasoning supervision and 13 behavioral patterns derived from real user interactions. The authors report a SOUL-Index of 65.7, above the strongest frontier model, and say it improves behavioral fidelity by 13.4 points and reduces Turing distance by 3.6 points versus the strongest baselines, Claude-Opus-5 on RealUserSim and SimulatorArena respectively. They then freeze the simulator to train an agent with multi-turn reinforcement learning, reporting better results than training with GPT-5.5 across nine unseen user simulators, and propose Coached On-Policy Self-Distillation (CSD) to turn the simulator's private reasoning traces into token-level supervision. Code is linked in the paper.
If independently verified, scalable user simulators could cut the cost of training and evaluating agents that normally depend on human feedback.