User simulators are meant to let AI assistants be trained and evaluated interactively at scale, but a new arXiv paper argues that these simulators have a fundamental shortcoming: they do not reproduce the diverse behaviors of real users. The abstract says this leads to assistants that generalize poorly, though the sentence is cut off in the source.

The paper, titled "Learning User Representations to Improve Simulation Fidelity," points to learning user representations as a way to close that gap. Because the abstract provided is truncated, the specific method and results are not available in the source, so the article should be read as a problem statement rather than a proven solution.

There is only one source here, so there are no conflicting findings to compare. The main takeaway is that simulation fidelity is seen as a bottleneck for building assistants that work well beyond their training scenarios.