Optimizing User Simulators via Conversation Reconstruction and Natural Language Representations

Researchers introduce a method to improve user simulator fidelity by generating natural-language descriptions that condition models to mimic specific user behaviors. The framework uses a conversation reconstruction task to evaluate representation quality by measuring how closely simulated interactions match original user data.