Persona-as-rationale distillation, live gallery — a 1.5B open language model generates a persona, a reasoning trace, and a 24-hour activity chain from coarse demographics alone. Trained by distilling a large open teacher's reasoning over real travel-survey days in Minneapolis–St. Paul (TBI 2023) and Seoul (KTDB 2021).
Demographics specify who a traveller is, but not how a person spends a day. The lifestyle signal that fills that gap is what we distill. A large open teacher model (Qwen2.5-32B) reads a real survey respondent's demographics together with that person's observed day and abduces the latent persona that best explains it, writing a forward-style chain-of-thought. A compact open student (Qwen2.5-1.5B + QLoRA) is then trained to reproduce the persona, the reasoning, and the activity chain from demographics alone. At deployment the student runs self-hosted with no teacher and no metered API in the loop — which is what makes a population-scale, auditable generator practical for public-sector planning.
On a three-group day-structure realism battery (per-person slot recovery, population-level distributional calibration, and behavioural plausibility), the distilled student leads classical structure-fitted baselines (marginal resampling, IPF, time-varying Markov chains, LSTM generators) on per-person fidelity in both cities, and honest training-side calibration carries it to the best overall battery score in both cities. Unlike the strongest classical baselines, which mostly replay day patterns seen in training, the student composes novel but realistic days.
Samples here are decoded at temperature 0.8 / top-p 0.92 from the Arm B student (persona + reasoning). The paper, code, and full evaluation are linked from the UMN Choi Lab site and github.com/UMN-Choi-Lab; a live in-browser (WebGPU) version of this demo runs the same model on your own GPU.