Beyond Raw Transcripts: Structured Persona Extraction for LLM-Based Digital Twins
Iris Ye, Tianze Deng, Ozan Candogan
Preprint. Submitted to NeurIPS 2026.
arXiv:2608.20344 [cs.CL]
Abstract
LLM-based "digital twins" aim to simulate how an individual would behave in new environments or respond to novel questions, given a representation of that individual's prior responses. A common approach constructs this representation from survey transcripts or summarized responses. Prior work shows that compressing long transcripts into shorter LLM-generated summaries does not significantly reduce predictive accuracy, suggesting that information volume is not the primary bottleneck.
In this work, we argue that the key limitation is instead structural: how persona information is organized before being provided to the simulator model. We study this by comparing unstructured summaries with structured persona representations. First, we introduce a hand-crafted schema (BDE: Background, Decision procedure, Evaluation), grounded in consumer-behavior theory, and show that it improves predictive accuracy over raw transcripts by +1.91 percentage points on a homogeneous benchmark (Twin-2K-500), with similar gains on gpt-5.4-mini and Qwen3-8B as robustness checks. However, this fixed structure does not generalize across more heterogeneous tasks, where performance is statistically indistinguishable from the raw transcript baseline.
To address this limitation, we propose an automatic structure-discovery pipeline in which an LLM iteratively proposes and refines task-specific persona structures and extraction prompts. On a benchmark of 13 diverse sub-studies, this approach restores performance, improving mean accuracy by +1.91 percentage points over the raw transcript baseline and eliminating significant losses observed with the fixed schema.
Overall, our results suggest that the main constraint in LLM-based digital twins is not how much information is provided, but how it is structured—and that the optimal structure depends on the task.
1. Introduction
Digital twins—virtual representations of individuals—have gained prominence as a tool for simulating human behavior in response to novel scenarios. These systems typically rely on large language models (LLMs) to generate predictions based on a person's historical responses, often captured in survey transcripts or summaries. As of 2026, the rapid advancement of LLMs has made such simulations increasingly feasible, yet the fundamental design choices for representing persona information remain underexplored.
Previous research has demonstrated that reducing long transcripts into concise, LLM-generated summaries does not meaningfully compromise predictive accuracy. This finding suggests that the volume of information is not the primary limiting factor. Instead, we hypothesize that the way persona information is structured exerts a more significant influence on downstream performance.
2. Problem Statement
We investigate whether structured persona representations outperform unstructured summaries in LLM-based digital twin simulations. Specifically, we ask: (1) Can a hand-crafted structure improve accuracy on homogeneous tasks? (2) Does such a fixed structure generalize across heterogeneous tasks? (3) Can automatic structure discovery adaptively optimize persona representation for diverse benchmarks?
3. Methodology
3.1 Hand-Crafted Schema: BDE
We introduce a schema grounded in consumer-behavior theory, consisting of three components:
- Background: Demographic and contextual information about the individual.
- Decision procedure: The individual's typical decision-making processes and heuristics.
- Evaluation: Criteria used to assess outcomes or options.
This structure is designed to organize raw transcript data into coherent categories, facilitating more effective use by the simulator model.
3.2 Automatic Structure Discovery
Recognizing the limitations of a fixed schema, we develop an iterative pipeline where an LLM proposes and refines task-specific persona structures and extraction prompts. This process adapts to the unique characteristics of each task, aiming to optimize information organization dynamically.
4. Experiments and Results
4.1 Homogeneous Benchmark
On the Twin-2K-500 dataset, the BDE schema improves predictive accuracy by +1.91 percentage points over raw transcripts, with consistent gains observed across gpt-5.4-mini and Qwen3-8B models, confirming robustness.
4.2 Heterogeneous Benchmarks
When applied to a benchmark of 13 diverse sub-studies, the fixed BDE schema yields performance statistically indistinguishable from the raw transcript baseline. This indicates that a one-size-fits-all structure is insufficient for varied tasks.
4.3 Automatic Structure Discovery Results
In contrast, our automatic structure-discovery pipeline achieves a +1.91 percentage point improvement in mean accuracy over the raw transcript baseline across the 13 sub-studies, effectively eliminating the significant losses associated with the fixed schema.
5. Discussion
Our findings underscore the importance of structural organization over information volume in LLM-based digital twins. While a hand-crafted schema can be effective for homogeneous tasks, optimal performance on heterogeneous benchmarks necessitates task-specific structuring. This suggests that future digital twin systems should incorporate adaptive mechanisms for persona representation.
6. Conclusion
We demonstrate that the primary constraint in LLM-based digital twins lies not in the amount of information provided, but in how that information is structured. Our work highlights the need for flexible, task-dependent persona extraction frameworks, with automatic structure discovery presenting a promising avenue for robust performance across diverse applications.
References
(References would follow here in a full paper.)
via ArXiv CL+LG