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

2026-05-28 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new adaptive interview framework addresses the challenge of accurately simulating individual decisions with large language models (LLMs) by gathering persona-relevant information through a structured three-stage dialogue. This process includes core questions, dynamic follow-ups, and a synthesized personality summary. The framework evaluates LLMs' ability to simulate participant decisions in moral dilemma scenarios, comparing performance across three conversational contexts: Core-10 responses, the full interview dialogue, and a summarized persona representation. Findings indicate that adaptive interviewing functions as a selective grounding mechanism rather than a uniform accuracy booster. Specifically, follow-up-derived evidence is incorporated in approximately 40% of full-interview traces, and these follow-up-grounded predictions achieve higher accuracy (45.5%) compared to core-only grounded ones (39.3%). This highlights that merely providing richer persona context is insufficient; improvements depend on models actively grounding their decisions in user-specific evidence.

Key takeaway

For NLP Engineers developing LLM-based persona simulations, relying solely on static descriptions will yield suboptimal results. Your models should integrate adaptive interviewing frameworks that gather dynamic, evidence-grounded persona information through multi-stage dialogues. This approach, which has shown to improve decision accuracy from 39.3% to 45.5% when evidence is grounded, ensures LLMs make more accurate, individually aligned predictions by actively using specific user evidence rather than just richer context.

Key insights

Adaptive interviewing improves LLM decision simulation by selectively grounding predictions in user-specific, follow-up derived evidence.

Principles

• Richer context alone is insufficient for LLM decision simulation.
• Decision accuracy improves when LLMs ground in specific user evidence.

Method

An adaptive interview framework uses a three-stage dialogue: core questions, dynamic follow-ups, and a synthesized personality summary to gather persona data for LLMs.

In practice

• Implement multi-stage persona data collection for LLMs.
• Prioritize evidence-grounding mechanisms in LLM decision tasks.

Topics

• Large Language Models
• Persona Simulation
• Adaptive Interviewing
• Evidence-Grounded Reasoning
• Decision Alignment
• Conversational AI

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Adaptive Querying with AI Persona Priors — AI personas from LLMs enable efficient, expressive Bayesian adaptive querying with closed-form posterior updates.
• ActTraitBench: Quantifying the Knowledge-Decision Gap in Large Language Models via Human-Grounded Behavioral Validation — ActTraitBench quantifies the Knowledge-Decision Gap in LLMs, revealing behavioral divergence despite consistent self-reports.
• Adaptive Thinking: Large Language Models Know When to Think in Latent Space — Self-consistency can predict thinking necessity, enabling adaptive allocation of LLM reasoning budgets.
• Teaching LLMs Brazilian Healthcare: Injecting Knowledge from Official Clinical Guidelines — Domain-specific synthetic data and reinforcement learning significantly enhance LLM performance on specialized clinical knowledge.
• I fine-tuned an LLM to be C-3PO to test which training data format works best for persona injection [P] — First-person statements are most effective for LLM persona injection, outperforming chat demos and synthetic docs.
• BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design — BED-LLM enhances LLMs' interactive information gathering via sequential Bayesian experimental design to maximize expected information gain.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.