Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent

Researchers introduce AgentMob, a training-free LLM-driven agent framework that improves mobility prediction by using adaptive evidence gathering rather than static prompts. The system achieves strong performance on multiple datasets by distinguishing routine cases from ambiguous ones, with significant accuracy improvements on difficult prediction scenarios.

AgentMob represents a meaningful advancement in how machine learning systems approach location prediction problems. Rather than relying on fixed prompts or single-pass inference, the framework implements a two-tier decision strategy: routine cases follow a fast path based on historical regularity, while ambiguous predictions trigger iterative tool use across multiple evidence sources including recent trajectories, behavioral history, and geographical factors. This mirrors how human decision-making operates under uncertainty—quick answers for familiar situations, deliberative reasoning for novel ones.

The research addresses a persistent limitation of current LLM applications: their tendency to commit to answers without seeking clarifying information. In mobility prediction, weak or conflicting signals frequently occur when users deviate from typical patterns. AgentMob's adaptive approach tackles this by allowing the system to request evidence dynamically rather than operating under a predetermined constraint. The training-free nature means practitioners can deploy the system immediately without costly fine-tuning cycles.

Performance metrics demonstrate tangible value: on the BW dataset's ambiguous cases, accuracy improved from 30.65% to 48.62% when using the LLM controller versus statistical baselines. This 60% relative improvement on difficult cases suggests the methodology effectively handles real-world prediction complexity. The framework's applicability extends beyond academic interest—urban planning, transportation optimization, and emergency response systems all depend on accurate mobility modeling.

The open-source release positions this work for broader adoption and refinement. Future developments might explore integration with real-time data streams, extension to multi-agent scenarios, or application to emerging use cases in autonomous systems and smart city infrastructure.

• →AgentMob uses adaptive evidence gathering instead of static prompts, improving accuracy on ambiguous mobility predictions by 60% over baselines.
• →The training-free framework operates with fast-path routing for routine cases and iterative tool use for uncertain scenarios.
• →Achieves 71.42% Acc@1 on BW dataset and 33-34% accuracy on larger real-world mobility datasets.
• →Open-source release enables immediate deployment without fine-tuning costs.
• →Demonstrates how LLMs can be enhanced through architecture design rather than scale alone.