Controllable User Simulation

Guy Tennenholtz
Ofer Meshi
Uri Shalit
Jihwan Jeong
Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS-26), Sydney (2026)

Abstract

Evaluating conversational agents with static offline datasets, while reliable, can be rigid, failing to cover rare scenarios or support testing new policies. This has induced increased reliance on user simulators for targeted, counterfactual evaluation. However, the theoretical foundations of these simulators remain understudied. In this work, we formalize controllable simulation as generation conditional on a random variable representing observable or latent targets of interest. By bridging natural language evaluation with classical off-policy evaluation (OPE) mathematics, we show that if controls are determined by full conversation trajectories, they couple to the data-generating agent policy, introducing a lookahead bias that breaks causal consistency. We also prove that under covariate shift, the variance of the trajectory density ratio explodes geometrically---a structural manifestation of the curse of horizon, which we term controllability collapse. We demonstrate mitigation strategies including strictly a priori controls, step-wise dynamic controls, and direct policy-conditioned learning. Empirical evaluation on WildChat and ConvApparel benchmarks show that static global controls distort conversational distributions and collapse behavioral diversity, while inference-time rejection sampling is computationally intractable. Conversely, our causally grounded simulators preserve natural variance and exhibit robust zero-shot generalization to unseen agent behaviors.

Research Areas

×