Training LLM agents to hold multi-turn conversations with users is more complex than single-turn tasks. The abstract for a new arXiv paper, EAU, argues that the learning signal for such agents is shaped jointly by the agent itself, the user simulator, and the environment. Because these three factors together determine what experiences the agent sees, changing any one of them changes what the agent can learn.

The authors propose co-evolving all three components during training. Rather than treating the user simulator and environment as fixed, EAU lets them adapt alongside the agent. This is meant to create a more realistic and responsive training loop, one that better reflects the give-and-take of real user interactions.

The paper is presented as a framework, and the abstract does not include experimental results. Still, the idea highlights a growing recognition that interactive agent training requires more than better agent architectures—it also needs better models of the people and contexts the agent will face.