← UMN Choi Lab

Activity Chain Generator

Pick a synthetic demographic profile to see a persona, a reasoning trace and an activity chain covering 24 hours in this precomputed gallery. The Arm B student, Qwen3.5-2B with a QLoRA adapter per region, generates all 3 from demographics alone for Minneapolis-Saint Paul, England, the Netherlands and Greater Melbourne.

What you are seeing. Profiles are synthetic, from a curated grid of gender, age band and employment status, plus home region for England. Remaining survey prompt fields, education, car licence, occupation and student status where available, are drawn from the training split conditional distribution within each cell, using only combinations shared by at least 5 training respondents. No real respondent record or observed day appears. Outputs are real, unedited generations precomputed on the lab GPU. Generations with invalid schemas are dropped; nothing else is filtered or cherry-picked.
Live version: The live page still runs the June 2026 build with the earlier Qwen2.5-1.5B students for Minneapolis-St. Paul and Seoul in the browser via WebGPU. The Qwen3.5-2B students are not yet compiled for the browser.
Loading samples…

How it works

The Qwen3.5-27B teacher reads the demographics of a survey respondent together with their observed day and abduces the latent persona that explains it, writing a reasoning trace in a forward style. The Qwen3.5-2B student, with a QLoRA adapter per region, is trained to reproduce the persona, reasoning and activity chain from demographics alone. The student is self-hosted at deployment, with no teacher or metered API in the loop. This makes an auditable generator at population scale practical for public-sector planning.

The 4 regions and surveys are Minneapolis-Saint Paul, Metropolitan Council Travel Behavior Inventory 2023; England, National Travel Survey 2023; the Netherlands, ODiN 2023; and Greater Melbourne, VISTA 2023-24. Synthetic profiles combine the curated grid with conditional draws from the training split; 3 samples per profile use decoding seeds 42 to 44. The student is compared with classical and neural baselines in all 4 regions using a realism battery for day structure with 3 groups: per-person slot recovery, population-level distributional calibration and behavioural plausibility. The paper reports the results.

Decoding uses temperature 0.8 and top-p 0.92. Project links: UMN Choi Lab and github.com/UMN-Choi-Lab. The live browser demo remains the June 2026 build.