LLM user simulators significantly affect agent task performance
UserProxyBench: Evaluating LLM User Simulators for Agent Benchmarks and Training
Artificial Intelligence
Summary
Interactive AI agents often rely on another AI to play the 'user' role, but current tests only check how well the agent does, not if the user AI followed instructions properly. The authors created UserProxyBench and a User Fidelity Score to measure how well the user AI sticks to its rules, separate from whether the agent succeeds. They found that changes in the user AI can change agent success scores, and often the user AI gives away information too early without affecting task success. This work helps pick the best cost-effective user simulators for testing and training AI agents.
What this means in practice
- •For ai development teams: Improve training and evaluation of interactive agents by measuring user simulator accuracy independently of agent success.
- •For enterprise ai integrators: Select cost-effective user simulators for agent benchmarks to better match enterprise task requirements.
Authors
Ashish Jain, Armaan Sandhu
Abstract
Interactive agent benchmarks and multi-turn reinforcement learning increasingly place a second language model in the role of the user. This simulated user controls what information the agent receives and when, yet current benchmarks score only the agent and do not directly measure whether the user correctly executed its assigned role. We introduce UserProxyBench, an evaluation layer over the tau-bench family, and the User Fidelity Score (UFS), which measures adherence to the benchmark's private user instructions using task-grounded rubric criteria scored independently of agent success. Holding the agent fixed at GPT-5.5 and varying only the user proxy across 375 enterprise tasks changes mean task reward by 15.2 points, while 24.4% of successful episodes contain a user-specification violation. The dominant failure is premature disclosure: users provide information before it is requested. This behavior has little effect on task reward, yet among successful episodes it causes the agent to make 1.06 fewer tool calls on average, changing the interaction being evaluated while preserving the reward. Finally, across seven proxies we identify an empirical cost-fidelity frontier, enabling practitioners to select the least expensive simulator that satisfies a required fidelity level.