User Simulation

HumanLM: Simulating Users with State Alignment Beats Response Imitation

🎉 Accepted at ICML 2026! Training user simulators to infer what real people believe and feel, not just mimic how they write, produces far more realistic simulated users.