Short answer: Organize the survey data before you hand it to the model. In a 2026 Chicago Booth study, LLM digital twins built from a structured profile (Background, Decision procedure, Evaluation) predicted people's held-out answers about 1.2 percentage points more accurately than twins given the raw survey transcript. Most of that gain came from the Evaluation profile and showed up in pricing questions.
Last updated: November 10, 2026
A digital twin is an LLM prompted with one real person's past answers so it can predict how that person would answer new questions. The usual design choice is how much of that history to feed the model. A new preprint, Structuring Survey Histories for LLM Digital Twins, by Iris Ye, Tianze Deng and Ozan Candogan of the University of Chicago Booth School of Business, asks a different question: does it matter how the history is organized? It is the camera-ready version for the SocialAgent workshop at NeurIPS 2026.
What did the study test?
The study tested whether a fixed, three-part profile predicts a person's answers better than their raw survey transcript. According to Ye, Deng and Candogan (2026), the authors used Twin-2K-500 (Toubia et al., 2025), a public benchmark in which 2,058 respondents answered 500 questions across three waves, with 88 questions from 17 tasks held out for evaluation.
Their representation, called BDE, splits each respondent's answers into three profiles:
- Background: who the person is and what they know and believe. This covers demographics, Big Five traits, values, political views and tested financial knowledge.
- Decision procedure: how the person reasons under uncertainty, drawn from need for cognition, need for closure, maximizing tendencies and error patterns on cognitive tests.
- Evaluation: what the person values, including budget constraints, risk and time preferences, social preferences and willingness to pay.
Each profile was written by gpt-4.1-mini at temperature 0 from the transcript, which never contained the held-out answers. Two simulators, gpt-5.4-nano and gpt-5.4-mini, then answered held-out questions as the persona. The evaluation used 100 respondents, each scored on 120 items (80 cognitive-bias items and 40 pricing items). Differences were tested with a paired bootstrap over respondents using 10,000 resamples.
Do structured profiles beat raw survey transcripts?
Yes, but modestly. The fully inferred BDE profile was 1.27 percentage points more accurate overall than the raw transcript on gpt-5.4-nano (p < 0.01) and 1.19 points more accurate on gpt-5.4-mini (p < 0.001). The clearest gains were in pricing, with +1.85 points on nano and +2.10 points on mini. Cognitive-bias accuracy improved significantly only on nano (+0.84 points).
The study does not show that shorter is better. The authors argue the gain comes from what the profile contains, not from compression. The study ran four contrasts to isolate the source of the gain:
| Contrast | What it isolates | gpt-5.4-nano (overall) | gpt-5.4-mini (overall) |
|---|---|---|---|
| BDE vs. raw transcript | Everything together | +1.27 pp (p < 0.01) | +1.19 pp (p < 0.001) |
| BDE vs. length-matched free-form profile | The three-part structure | +1.05 pp (p < 0.05) | +0.57 pp (not significant) |
| Deeper vs. shallow inference during extraction | How much the extractor may infer | +0.06 to +0.83 pp (not significant) | +0.11 to +0.27 pp (not significant) |
| Adding Evaluation to Background + Decision procedure | Preference content | +1.44 pp (p < 0.001) | +1.22 pp (p < 0.001) |
Source: Ye, Deng & Candogan (2026), Tables 1 to 4. pp = percentage points of accuracy, n = 100 respondents.
Which part of the profile matters most?
The Evaluation profile matters most. Adding it to Background plus Decision procedure raised pricing accuracy by 3.40 points on nano and 2.60 points on mini (both p < 0.001). Adding the Decision procedure to Background plus Evaluation produced no significant gain on any metric.
This is a useful result for anyone building twins for product or pricing research. Whether someone would buy a given grocery item at a given price depends on their budget, risk attitude and willingness to pay more than on their reasoning style. A profile that surfaces those preferences explicitly makes it easier for the simulator to use them.
Does deeper inference or formatting help?
No. Letting the extractor infer more aggressively, from labeled demographic inference up to broad inference and a reliability assessment, produced no significant improvement on either simulator. Removing the section headers from the BDE text also made no significant difference. What helped was the content each section was designed to pull out, not the markdown labels.
The free-form comparison supports the same point. A free-form profile written by the same model, under the same inference rules and matched to within 10% of BDE's length, was still less accurate on pricing: 2.15 points lower on nano (p < 0.01) and 1.38 points lower on mini (p < 0.05).
How much should researchers trust these results?
Treat them as a well-controlled but narrow result. The authors list four limitations themselves:
- One respondent panel, one benchmark and two models from a single model family.
- The free-form comparison changes what is extracted and how it is organized at the same time, so it cannot separate the two.
- The Evaluation ablation has no equal-length control, so preference content and extra relevant information are confounded.
- The overall score weights pricing as a third of all items. When each of the 17 tasks is weighted equally, only the end-to-end overall advantage on gpt-5.4-mini stays significant. The pricing gains are unaffected.
The effect sizes are also small, at one to three percentage points. The paper does not claim the three-part split is the best possible design or that the profiles capture real psychology.
What this means for researchers using digital twins
The study's practical lesson is to curate the twin's input around what your study needs to predict. Three takeaways follow:
- Do not assume the full transcript is the best input. In this study a structured profile beat the raw history on both simulators.
- Put preference data up front for commercial questions. Budget, risk, time preferences and willingness to pay drove the pricing gains.
- Validate against held-out answers. The design here (paired comparisons on held-out questions, bootstrap intervals) is a template for checking any twin before relying on it. Earlier work on individual-level digital twin accuracy shows why that check matters.
If you run twin studies yourself, the Qualitati Digital Twin Panel builds personas from real participant data, and the same discipline applies there: test the twin against answers it has not seen before you use its predictions.
FAQ
What is an LLM digital twin?
It is a language model given one real person's prior responses, such as survey answers, and asked to predict how that person would respond to new questions. It differs from a generic synthetic persona because it is anchored to an individual's own data.
What is the BDE profile?
BDE is a fixed schema from Ye, Deng and Candogan (2026). It splits a respondent's survey history into Background (identity, knowledge, beliefs), Decision procedure (how they reason under uncertainty) and Evaluation (what they value, including budget, risk and willingness to pay).
How big was the accuracy gain?
About 1.2 to 1.3 percentage points overall versus the raw transcript, and up to 3.40 points on pricing when the Evaluation profile was added. The gains are statistically significant but small.
Is the paper peer-reviewed?
It is an arXiv preprint (version 1, 5 November 2026), described as the camera-ready version for a NeurIPS 2026 workshop. Workshop papers are reviewed, but less extensively than journal articles.
Source: Ye, I., Deng, T., & Candogan, O. (2026). Structuring Survey Histories for LLM Digital Twins: A Controlled Study of Background, Decision Procedure, and Evaluation Profiles. arXiv:2611.08730. SocialAgent Workshop, NeurIPS 2026.
Last updated: November 10, 2026. This article is an independent editorial summary of third-party research by the Qualitati Research Team. It is not affiliated with or endorsed by the paper's authors.