LLM agents can improve prediction accuracy on ambiguous cases by dynamically deciding what evidence to gather, rather than using static prompts—showing a practical way to combine LLM reasoning with structured decision-making.
This paper presents AgentMob, an LLM-based agent that predicts where people will go next by adaptively gathering evidence rather than relying on fixed prompts. It uses a fast path for routine movements and triggers iterative tool use (checking historical patterns, geography, movement likelihood) when predictions are uncertain, achieving strong results without task-specific training.