Beyond Individual Personas: Aligning Synthetic Dialogue to Population-Level Behavior Distributions

2026-06-05

AI Summary

This work addresses a critical limitation in existing persona-based dialogue generation methods, which prioritize single-turn response quality while overlooking discrepancies between synthetic and target population-level behavioral distributions. To bridge this gap, we propose GroupPersona, a novel framework that explicitly aligns group-level behavioral distributions during dialogue generation. Our approach models core behavioral attributes, extracts population-level statistical features from reference corpora, and leverages these as conditional control signals to guide the synthesis process—thereby preserving dialogue structure while modulating agent interaction patterns. Experimental results across four datasets demonstrate that GroupPersona reduces the average Jensen–Shannon divergence across twelve behavioral attributes between synthetic and reference corpora by 24.4%, while significantly improving the calibration accuracy of dialogue quality metrics.

Abstract

Synthetic dialogue corpora are increasingly used as proxies for target dialogue data, yet persona-grounded generators optimize individual conversations rather than corpus composition, yielding locally plausible dialogues with distorted population-level behavior mixes. We introduce GroupPersona, a framework that aligns synthetic dialogue corpora to the behavior distribution of a reference corpus. GroupPersona turns population statistics into generation controls: it separates each dialogue's core behavioral signature from predictable side effects, and uses the resulting behavioral groups to condition user agents on the interaction patterns that define the reference population. We evaluate GroupPersona on four corpora crossing two dialogue sources, assistant-style and Reddit-derived, with two construction variants: structure-preserving and variation-enhanced. GroupPersona lowers Jensen-Shannon divergence between synthetic and reference distributions over 12 behavior attributes from 0.234 to 0.177 relative to the strongest average baseline, a 24.4% reduction, and is best or tied-best on all four corpora while preserving structural alignment. It also achieves the closest calibration to reference-conversation quality scores, reducing mean absolute deviation from the reference-conversation profile to 0.63 versus 0.91 for the next-best baseline.

Problem

• synthetic dialogue
• population-level behavior
• behavior distribution
• persona-grounded generation
• corpus alignment

Innovation

• synthetic dialogue
• population-level alignment
• behavior distribution
• GroupPersona
• persona grounding