Using Base Models for More Natural User Simulations

_ddjohnson · x · 2026-09-01

To make simulated users more realistic, the system uses a pretraining-only base model to generate user messages, with an assistant-trained "pilot" selecting the best candidate. This approach increases naturalness without compromising controllability.

Related event: OpenAI Uses Simulated Users for Multi-Turn LLM Evaluation(2 posts)→

Original post →

More from Research

Research channel →