DIAL: Direct Iterative Adversarial Learning for Realistic Multi-Turn Dialogue Simulation

TL;DR

DIAL is an adversarial learning framework that improves the realism of multi-turn dialogue user simulators, especially in sensitive domains like mental health support, by enhancing diversity and failure detection capabilities.

Contribution

Introduces DIAL, a novel adversarial approach that iteratively refines user simulators to better mimic human behavior and identify system failures.

Findings

Restores lexical diversity in simulated dialogues.

Reduces discriminator accuracy, indicating more realistic simulation.

Strong correlation between simulated and real failure rates.

Abstract

Realistic user simulation is crucial for training and evaluating multi-turn dialogue systems, yet creating simulators that accurately replicate human behavior remains a significant challenge. An effective simulator must expose the failure modes of the systems under evaluation. This work introduces Direct Iterative Adversarial Learning (DIAL), an adversarial framework that iteratively enhances user simulator realism through a competitive dynamic between a generator (user simulator) and a discriminator. When applied to mental health support, a domain characterized by diverse failure types and a critical dependence on realistic user behavior for failure detection, DIAL restores lexical diversity diminished by supervised fine-tuning and drastically reduces discriminator accuracy. The resulting simulator exhibits a strong correlation between simulated and real failure occurrence rates while maintaining low distributional divergence of failure modes. These findings indicate that DIAL is a promising method for developing realistic user simulators in multi-turn dialogue, facilitating reliable and cost-effective system evaluation prior to deployment.