UserBench was jointly proposed in July 2025 by the Salesforce AI Research team and a research team from the University of Illinois at Urbana-Champaign. The related research results were published in the paper "...".UserBench: An Interactive Gym Environment for User-Centric Agents".

UserBench is a user-centric benchmark designed to evaluate the performance of agents in multi-turn, preference-driven interactions. In UserBench, simulated users provide initial, vague task instructions, gradually revealing preferences over time, often implicitly. Agents must proactively clarify their goals, interpret subtle cues, and succeed through adaptive reasoning tools. Built on the standard Gymnasium framework, UserBench offers a modular, scalable setup with standardized interaction interfaces and a stable backend for tool usage, enabling rigorous and repeatable evaluation.