How Does a Language Model Learn Crowd Behavior From Only Aggregate Counts?
Two researchers fine-tuned a language model to simulate where people go after a baseball game. The model learns from mobile network counts instead of individual tracks.
The data problem starts with privacy
Pedestrian simulators need a rule for every agent. Privacy limits the input to zone-level device counts and origin-to-destination flows. No individual trajectories exist. Those aggregates under-determine behavior because many different decision sets reproduce the same counts. The paper frames this as a distillation task where the model must infer individual choices from collective totals.
The target comes from the OD flow
The team reads the destination composition from the flow data. That composition is the fraction of the departing crowd heading to each point of interest. They reweight the model's own destination distribution onto that target using iterative proportional fitting. The fitting happens before training.
Fine-tuning inflates the dominant class
A low-rank adapter is fitted to trajectories resampled to a corrected training composition. The correction anticipates the inflation that fine-tuning causes. After inflation the simulated population reaches the target. The agent then runs without inference-time correction.
Two baseball games supply the evidence
Mobile network counts from two games provide the test bed. The fine-tuned agent cuts destination-share error by 25%. Grid correlation stays similar across policies.
Where the explanation stops
The abstract does not say how many agents were simulated. It does not name the language model base. It does not describe the resampling algorithm beyond the corrected composition. It does not report whether the 25% reduction is statistically significant. The full text may answer some of these. The abstract does not.
The sources this article was written from
- "arXiv:2608.19778v1
https://arxiv.org/abs/2608.19778v1"