Adaptive Interviewing for Persona Simulation in LLMs: Evidence-Grounded Reasoning Improves Decision Alignment

Researchers propose an adaptive interview framework to improve how large language models simulate individual decision-making by gathering persona-relevant information through structured dialogue. The study finds that richer contextual information alone doesn't guarantee better accuracy; instead, LLMs only improve predictions (45.5% vs. 39.3%) when they actively ground decisions in user-specific evidence extracted during follow-up questions.

This research addresses a fundamental limitation in LLM capabilities: the difficulty of accurately modeling how specific individuals make decisions. Traditional approaches provide static persona descriptions that fail to capture the nuanced values, experiences, and contextual cues that drive human decision-making. The adaptive interview framework represents a meaningful shift toward dynamic, evidence-based persona construction through three-stage dialogue that progressively refines understanding of an individual's decision-making patterns.

The findings challenge a common assumption in AI development—that more information automatically produces better outcomes. Instead, the research reveals a selective grounding mechanism where follow-up questions generate valuable evidence in only 40% of cases, yet these evidence-grounded predictions significantly outperform baseline accuracy. This distinction matters because it suggests the quality and relevance of information matters more than quantity, and models must actively incorporate specific evidence into their reasoning chains to achieve improvements.

For developers building AI applications requiring personalized decision simulation—from healthcare advisory systems to financial planning tools—this research demonstrates that effective persona modeling requires interactive, adaptive data gathering rather than static profiles. The methodology could inform how AI systems conduct user interviews, onboarding processes, and preference elicitation workflows.

Looking forward, this work opens questions about how to systematically identify which types of follow-up questions generate actionable evidence for specific decision contexts, and whether the 40% incorporation rate reflects model limitations or optimal filtering of relevant information. Future research should explore scaling this framework across diverse persona types and decision domains to establish broader applicability.

• →Adaptive interviewing improves LLM decision simulation accuracy to 45.5% when grounded in follow-up evidence versus 39.3% for core-only responses.
• →Evidence incorporation occurs in approximately 40% of full-interview reasoning traces, indicating selective rather than comprehensive information use.
• →Richer persona context alone fails to improve predictions; models must actively ground decisions in user-specific evidence to achieve gains.
• →The framework uses three-stage dialogue—core questions, dynamic follow-ups, and synthesized summaries—to extract decision-relevant information.
• →This research suggests interactive data gathering outperforms static persona descriptions for individual-level decision simulation in LLMs.