Controllable User Simulation

TL;DR

This paper formalizes controllable user simulation as a causal inference problem, identifying biases in existing methods and proposing mitigations to improve evaluation of conversational agents.

Contribution

It introduces a causal framework for controllable simulation, analyzes biases in supervised fine-tuning, and proposes practical training methods for unbiased, robust simulators.

Findings

Standard fine-tuning introduces look-ahead bias and reduces diversity.

Causally grounded simulators eliminate bias and maintain natural variance.

Proposed methods improve zero-shot generalization to unseen behaviors.

Abstract

Using offline datasets to evaluate conversational agents often fails to cover rare scenarios or to support testing new policies. This has motivated the use of controllable user simulators for targeted, counterfactual evaluation, typically implemented by prompting or fine-tuning large language models. In this work, we formalize controllable simulation as a causal inference problem. By bridging natural language evaluation with off-policy evaluation methodology, we show that the standard practice of training simulators via supervised fine-tuning on post-hoc trajectory labels yields a structurally biased model. Specifically, these labels are inextricably coupled to the data-generating behavior policy, injecting a look-ahead bias that breaks causal consistency. Furthermore, we prove that under policy shift this failure causes the variance of evaluation metrics to explode geometrically, a phenomenon we term controllability collapse. To restore causal consistency, we establish theoretical conditions for accurate simulation and propose practical training mitigations: a priori controls, step-wise dynamic controls, and direct policy-conditioned learning. Empirical evaluation confirms that while standard global controls distort conversational distributions and collapse behavioral diversity, our causally grounded simulators eliminate look-ahead bias, preserve natural variance, and exhibit robust zero-shot generalization to unseen agent behaviors.