MobEvolve Rewrites Its Own Rules to Beat Deep AI Mobility Models
MobEvolve uses an LLM agent to iteratively rewrite its own heuristic rules, beating deep generative models at synthetic trip-chain generation while staying fast and interpretable
Takeaways
- MobEvolve is a new framework that uses an LLM agent to iteratively rewrite explicit heuristic rules for generating synthetic human trip chains.
- It outperforms deep generative models and LLM-based methods on individual fidelity, population distribution, and behavioral plausibility on Singapore and Montreal benchmarks.
- The key insight: the LLM is only active during an evolution/training phase; inference runs on the evolved rules alone, making it fast and scalable.
- The heuristic pipeline covers travel participation, destination retrieval, individual adaptation, and behavioral refinement -- all as auditable, human-readable rules.
- An evolution memory accumulates successful rule edits across iterations, enabling cumulative improvement without gradient-based training.
- No code has been released yet; the paper is available on arXiv.
Generating realistic synthetic human mobility data is one of those problems that sounds simple but hides enormous complexity. You need a system that can reproduce where a person goes, in what order, for how long, and why -- all while matching population-level statistics, respecting individual behavioral quirks, and remaining fast enough to scale to millions of people. A new paper introduces MobEvolve, a framework that takes a radically different approach: start with a transparent, rule-based heuristic system, then let an LLM agent rewrite those rules until they work.
The result is a system that outperforms state-of-the-art deep generative models and LLM-based methods on two real-world benchmarks, while being both interpretable and computationally efficient -- a combination that has so far eluded the field.
Why this problem is harder than it looks
Human mobility generation aims to synthesize plausible spatio-temporal trajectories of human movement, and the study of such trajectories offers deep insights for urban planning, transportation management, and public health. The practical demand is real: mobility data is typically derived from user tracking and travel surveys, which are subject to privacy concerns, participant noncompliance, and high cost. Synthetic data sidesteps all of that.
The field has tried three main approaches, each with a fatal flaw:
- Traditional activity-based models (ABMs): These rely on rigid assumptions and require extensive data, making them costly and difficult to adapt to new regions, especially those with limited conventional travel data.
- Deep generative models (VAEs, GANs, diffusion): The generated data only imitates real-world data from the data distribution perspective rather than semantics, rendering them less effective in simulating or interpreting activities in novel or unforeseen scenarios with a significantly different distribution.
- LLM-based agents: Simulating large-scale LLM agents remains highly challenging. Running a full LLM inference call per person per trip is simply not practical at city scale.

The core tension is that interpretability and efficiency pull in opposite directions from accuracy. Deep models can fit distributions well but are black boxes. LLMs reason semantically but are slow. Heuristics are fast and transparent but hard to tune. MobEvolve's bet is that you can get all four properties at once -- if you let an LLM do the tuning automatically.