Skip to main content
Back

AgentMob Beats Every Training-Free LLM at Predicting Where You Go Next

AgentMob skips LLM calls for routine trips and only invokes iterative tool use when location signals are genuinely ambiguous, hitting 71% top-1 accuracy on GPS data

Paper, Full Access
  • AgentMob is a training-free LLM agent for next-location prediction that uses a fast path for routine trips and an iterative tool-calling slow path for ambiguous cases.
  • On hard non-routine cases, the LLM controller boosts top-1 accuracy from 30.65% to 48.62% over a same-tool statistical baseline on the BW dataset.
  • Overall best training-free LLM performance across three datasets: 71.42% on BW, 33.14% on YJMob100K, and 33.50% on Shanghai ISP using GPT-4.5.
  • Four tools are available to the agent: Mobility Context Retriever, Historical Location Retriever, Daily-Move Estimator, and Geographical Information Retriever -- called adaptively, up to 10 iterations.
  • Unlike prior agentic methods that always run all modules, AgentMob stops gathering evidence as soon as the LLM is confident, saving compute on easy predictions.
  • Code is open-source on GitHub; the framework is model-agnostic and requires no task-specific training data.

Predicting where someone will go next sounds deceptively simple. In practice, it is one of the harder problems in urban AI: human movement is mostly routine, but just irregular enough that the edge cases break most models. A new paper introduces AgentMob, a training-free LLM-driven agent framework that tackles this by treating next-location prediction as an adaptive decision process rather than a fixed inference pipeline.

A field caught between accuracy and explainability

Urban mobility prediction focuses on modeling and forecasting individual movements within cities based on historical patterns and factors such as time, urban structure, and activity purposes, and is essential for applications such as traffic management and location-based services. The stakes are real: better predictions feed into transit optimization, epidemic modeling, and urban simulation at scale.

The field has gone through two generations of approaches. With the rise of deep learning, sequence modeling approaches became prominent, represented by RNN-based methods such as DeepMove and LSTPM. These supervised models are accurate but brittle: they need task-specific training data, do not generalize across cities, and offer almost no insight into why they made a prediction. A shared motivation for LLM-based alternatives is that mobility datasets often contain heterogeneous and weakly structured information -- like POI semantics and trip context -- and LLMs are attractive because they can exploit semantic knowledge about places and activities and support few-shot or zero-shot prediction when conventional models struggle to generalize.

Recent research leverages LLMs as predictive models for urban mobility tasks using prompt engineering and few/zero-shot learning, with studies demonstrating LLMs' ability to predict individuals' next visits without city-specific training. But these approaches have a critical flaw: they rely on static prompts and single-pass inference. When the mobility signal is weak -- say, a user deviates from their usual Tuesday routine -- the model has no mechanism to go look for more evidence. It just guesses.