BDE Structured Profiles Improve LLM Digital Twin Predictions
Summary:This paper studies BDE, a fixed representation that decomposes an individual's survey history into Background, Decision procedure, and Evaluation before an LLM predicts how that person would answer new questions. On 17 held-out tasks from Twin-2K-500, complete BDE profiles improve overall accuracy over raw transcripts by 1.27 percentage points with gpt-5.4-nano and 1.19 points with gpt-5.4-mini. Comparisons with length-matched free-form profiles suggest that the benefit comes primarily from prof
Core Finding
LLM-based digital twins typically predict how an individual would answer new questions by using that person's earlier survey responses. This study shows that performance depends not only on how much information is retained, but also on how it is represented. The proposed BDE structure divides a survey history into Background, Decision procedure, and Evaluation, providing fixed semantic slots for stable facts, the process used to make choices, and the criteria used to assess outcomes.
On Twin-2K-500, the authors evaluated held-out predictions across 17 tasks. In a paired evaluation with 100 respondents, complete BDE profiles improved overall accuracy over raw survey transcripts by 1.27 percentage points with gpt-5.4-nano and by 1.19 points with gpt-5.4-mini. The reported significance levels were p < 0.01 and p < 0.001, respectively.
How BDE Works
BDE is neither a verbatim recap nor a generic summary. It constrains profile construction to a fixed set of semantic components. A practical pipeline can extract stable background facts, reconstruct the respondent's decision rules and trade-offs, summarize evaluation criteria and feedback from prior choices, and then combine these components into a reusable profile for predicting new responses.
The fixed schema offers practical advantages over free-form summaries. Downstream components do not need to infer whether a sentence represents a stable preference or a context-specific reaction. Profiles can also be versioned, audited, updated incrementally, and tested through component-level ablations.
The approach still depends on LLM-based extraction. It reduces ambiguity and redundancy in long transcripts, but it does not eliminate factual errors. Invented background facts, unsupported decision rules, or overconfident psychological inferences can be persisted into the profile and then propagated into predictions.
Why Compression Alone Does Not Explain the Gain
The study includes a particularly informative baseline: the same extractor reads the same transcript and produces a length-matched free-form profile. BDE improves pricing-task accuracy by 2.15 percentage points with nano and 1.38 points with mini. Its overall advantage reaches statistical significance only for nano, where the gain is 1.05 points.
This comparison argues against a simple explanation that shorter or cleaner prompts are responsible for the improvement. Part of the gain appears to come from structuring decision-related semantics. It also reveals a limitation: BDE is not uniformly superior across all tasks and model variants, and the mini model's overall result does not show a statistically significant advantage.
Profile Composition Matters More Than Deeper Extraction
Component ablations show that adding Evaluation to a profile containing only Background and Decision procedure raises overall accuracy by 1.44 points on gpt-5.4-nano and 1.22 points on gpt-5.4-mini. Both gains are highly significant. Evaluation apparently supplies information about the respondent's standards and prior feedback that is difficult for the model to extract directly from a raw transcript.
By contrast, allowing the extractor to perform deeper inference does not produce a statistically significant gain on either simulator. Longer reasoning chains and more expensive profile generation therefore do not automatically improve downstream predictions. The more useful engineering question is which information should be represented explicitly and whether a component reflects a stable decision mechanism, rather than simply how much inference budget is spent.
Engineering Trade-offs and Evaluation Limits
BDE shifts work toward offline profile extraction. The pipeline adds preprocessing cost but can reduce redundant context during repeated prediction, enable profile caching, and support component-level auditing. It also introduces a new failure mode: an extraction mistake can be frozen into a reusable artifact. A newly added component does not guarantee improvements on every downstream task.
The evaluation is also limited. The main paired test covers 100 respondents, and the outcome can depend on the survey population, task distribution, and question design. Accuracy is the dominant metric, while calibration, cross-domain robustness, profile length, extraction cost, and human-rated profile validity are not sufficiently characterized. The difference between nano and mini further indicates that structured prompts do not remove sensitivity to the model's parsing and reasoning capabilities.
Digital twins also create privacy concerns. Converting survey transcripts into structured profiles creates another sensitive dataset. Systems should restrict intermediate storage, retain evidence links for each extracted field, version every profile, and avoid turning one-time answers into durable psychological claims.
Deployment Guidance
Developers can adopt BDE incrementally as a profiling template rather than committing to the entire pipeline immediately. Preserve source evidence for each component, run paired tests with raw transcripts, and treat Evaluation as a high-priority addition while validating it per task. Measure extraction cost, cache-hit rate, profile-update frequency, and sampled factual errors. Most importantly, compare structured profiles against both raw context and length-matched summaries; otherwise, it is impossible to determine whether the gain comes from structure, compression, or the model itself.
BDE's broader lesson is that a digital-twin profile should explicitly represent how an individual decides and evaluates outcomes. For engineering teams, a carefully designed schema may deliver more value than adding ever longer context windows or deeper inference traces.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-11-09)
Tags: #LLM Digital Twins #Structured Profiling #Survey Prediction #Twin-2K-500
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments