The race to LLM-based digital twins has so far followed a seemingly obvious premise: the more information you feed the model, the more accurate its simulation of a person will be. A new study turns that assumption around. The bottleneck, the researchers argue, is not the amount of data but the structure in which personal information is organized before being passed to the simulator model.

The study compares two approaches. On one side, unstructured summaries generated by LLMs from survey transcripts. On the other, structured representations of the person. On a homogeneous benchmark called Twin-2K-500, a hand-crafted schema — BDE, standing for Background, Decision procedure, Evaluation — improves predictive accuracy by +1.91 percentage points over raw transcripts. The gain holds with gpt-5.4-mini and Qwen3-8B, two models that differ significantly in architecture and access mode.

But there is a limit. The fixed BDE schema does not hold up when tasks become heterogeneous: in that case, performance returns to being statistically indistinguishable from the baseline. To overcome this, the researchers propose an automatic structure-discovery pipeline in which an LLM iteratively proposes and refines task-specific schemas and extraction prompts. On a benchmark of 13 diverse sub-studies, this approach restores the advantage, with a mean improvement of +1.91 percentage points over the baseline and the elimination of the significant losses seen with the fixed schema.

The message is clear: for digital twins, the main constraint is not how much information is provided, but how it is organized. And the optimal structure depends on the task.

This conclusion has direct implications for those designing inference pipelines on self-hosted infrastructure. In a local setting, context window and available VRAM are finite resources. If compressing transcripts into summaries does not reduce accuracy but structure does, then the work of designing extraction schemas takes on greater weight than accumulating raw tokens. A structured representation can reduce the input to the simulator and ease memory pressure without sacrificing fidelity. This is not a universal recommendation, but a trade-off to evaluate: the right schema can be cheaper in token terms, but the automatic search for structure introduces additional inference steps of its own.

There is also a data sovereignty angle. Structured extraction can happen entirely locally, limiting the need to send raw transcripts to cloud services. This aligns with the data minimization principles of GDPR. Moreover, the fact that the results were also verified with Qwen3-8B — an open-weight model that can run on one's own hardware — signals that the approach does not depend on proprietary APIs.

From a competitive standpoint, the finding reduces the value of generic summarization tools. If volume is not the problem, vendors selling compression alone risk answering the wrong question. Conversely, developers of orchestration frameworks for dynamic prompts and schemas may find room. Structure becomes a design layer of its own, distinct from model fine-tuning and data collection.

Ultimately, the work shifts attention from a race for content to a race for form. This is not an academic detail: for those building digital twins on local stacks, it means rethinking pipelines not as simple data conduits, but as systems that must continuously negotiate between fidelity, compute costs, and context constraints.