Grouped Simulator: Bridging Group Homogeneity and Individual Heterogeneity for High-Fidelity User Simulation

ACL ARR 2026 January Submission198 Authors

Keywords: Conversational AI

Abstract:

High-fidelity user simulation is critical for optimizing downstream multi-turn conversational applications such as telemarketing and automated customer service.
Current approaches based on Large Language Models typically employ role-playing via prompting, relying on coarse-grained profiles and internal knowledge to guide behavior generation.
However, a realistic simulation requires the simultaneous modeling of group homogeneity and individual heterogeneity. The former refers to shared fine-grained behavioral patterns within a user cohort, while the latter represents the diverse preferences and expression styles of distinct users.
Existing paradigms struggle to meet this dual requirement, failing to capture nuanced group commonalities while being insufficient in personalized diversity.
To address these limitations, we propose the Grouped Simulator, a framework designed to bridge group homogeneity and individual heterogeneity.
We implement a dual-optimization strategy: (1) Group-Aligned Reinforcement Learning with multi-level reward to internalize shared behavioral patterns and linguistic norms, and (2) a Retrieval-Augmented Dynamic SOP Engine to inject diverse, context-aware individual feedback.
Extensive experiments in telemarketing scenarios demonstrate that Grouped Simulator significantly outperforms state-of-the-art baselines in terms of realism and diversity.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Dialogue and Interactive Systems

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 198