Training interactive language agents typically requires rich user interactions, but gathering human feedback is costly and does not scale well. Simulated users offer a promising alternative, yet they need to behave realistically enough to provide useful training signal. A new paper on arXiv introduces MIMESIS, a method for learning user simulators as training environments for such agents.
The core idea is to replace expensive human interactions with learned simulators that can generate user behavior. This allows agents to be trained and evaluated at scale, without constant human involvement. However, the paper notes that the effectiveness of this approach hinges on the realism of the simulated users—if they diverge too far from real human behavior, the trained agents may not generalize.
MIMESIS appears to address this by learning simulators from data, rather than hand-crafting rules. While the abstract is truncated, the emphasis on learning suggests a data-driven approach to capturing user patterns. The paper positions this as a step toward scalable training environments for interactive agents, though it does not yet report experimental results in the provided text.