Synthetic user profiles have a structural flaw: they flatten individual variation, draw on stereotypes, and miss the signals that drive real preferences. A ServiceNow team has tried to reverse this with a framework called profile behavioral grounding, which extracts open-ended, high-fidelity user profiles directly from authentic, anonymized social media posts. The code is public on GitHub.

The novelty is not another persona dataset but the construction method. Instead of defining a demographic identity from the start, the system starts from observed behaviors and turns them into profiles usable in two different regimes: supervised fine-tuning (SFT) for train-time personalization and non-parametric multi-perspective reasoning at test time. On recommendation and open-ended query benchmarks, the behaviorally grounded profiles improve base models and outperform synthetic baselines.

There is an underlying thesis: LLM personalization is shifting its center of gravity from identity templates to behavioral signals. This has direct implications for teams managing self-hosted models. A company that fine-tunes a model on its own data cannot just invent characters: it must build extraction pipelines that turn real conversations, tickets, and interactions into usable profiles while maintaining anonymization and data control. This shift is not neutral for infrastructure: processing real textual data requires more elaborate cleaning, filtering, and token handling pipelines than using already ordered synthetic datasets.

The winners are organizations with access to large volumes of authentic interactions: they can obtain models more aligned with real behavior without relying on pre-packaged profiles. The losers are providers of synthetic persona datasets, whose added value shrinks if teams can extract profiles from their own data flows. But there is also a governance cost: real data, even anonymized, carries risks of bias and exposure of sensitive information, and profile quality depends on the ability to filter noise and unrepresentative content.

Structurally, the work signals that behavior-based personalization can become a differentiator for organizations operating in on-premise or hybrid environments, where data must not leave the corporate perimeter. For those evaluating self-hosted deployments, AI-RADAR offers on /llm-onpremise some analytical frameworks to assess trade-offs between data control, pipeline costs, and profile quality. The question is not whether synthetic personas will disappear, but whether extraction processes from real data will become scalable enough to be a credible alternative in personalization workflows. On this study, the direction is already set.