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

Technical Summary: Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agents

Problem Definition
Individual-level mobility prediction is critical for urban simulation, transportation planning, and public health analysis. Existing approaches generally fall into two paradigms, both with significant limitations:

1. Supervised Sequence Models: Methods like RNNs and Transformers achieve high accuracy by learning spatiotemporal dependencies from large-scale data. However, they are "black boxes" with limited decision-level transparency, require task-specific training (making adaptation to new cities or granularities costly), and cannot dynamically seek additional information when signals are weak.
2. Static LLM-Based Methods: Recent approaches convert trajectories into text and leverage Large Language Models (LLMs) for interpretability. However, most rely on static prompts and single-pass inference. They compress all available evidence into a fixed context, lacking the ability to iteratively retrieve, cross-check, or revise predictions when mobility signals are conflicting or ambiguous.

The core challenge addressed is that mobility instances vary in difficulty. Routine cases (e.g., a user visiting the same location at a specific time) are easily solved by historical regularity, while ambiguous cases require reconciling conflicting signals from recent context, long-term habits, stay-move tendencies, and geographical plausibility. Current methods fail to adapt their reasoning effort to these varying difficulties.

Methodology: AgentMob
The authors propose AgentMob, a training-free, LLM-driven agent framework that reframes next-location prediction as an adaptive evidence-controlled decision-making process. Instead of treating the LLM as a direct predictor, AgentMob uses it as a controller that dynamically decides how much evidence is required for a specific instance.

The framework operates through the following pipeline:

1. Input Processing: Raw GPS records are discretized into spatial units (administrative polygons or grid cells) and paired with textual descriptions of urban functions (POIs, region descriptions).
2. Fast-Path Prediction: The agent first performs a lightweight check for historical regularity. If a user's history shows a dominant location for the specific weekday and hour, the agent immediately returns this location. This bypasses expensive computation for routine cases.
3. Adaptive Tool-Calling (for Ambiguous Cases): When historical regularity is insufficient, the agent enters an iterative loop (capped at 10 iterations) to gather evidence. It invokes a Mobility Analysis Toolbox containing four distinct tools:
   • Mobility Context Retriever: Summarizes recent trajectory sequences and same-time visitation statistics to distinguish immediate continuity from long-term patterns.
   • Geographical Information Retriever: Provides distance metrics and textual descriptions of candidate locations to verify geographical plausibility and semantic consistency.
   • Stay–Move Estimator: Estimates the likelihood of the user remaining at the current location versus moving, based on current stay duration versus historical statistics.
   • Historical Behavior Retriever: Retrieves behavioral patterns (dwell times, frequent next destinations) from past visits to similar locations under comparable temporal contexts.
4. Decision and Auditability: The LLM controller synthesizes the evidence from these tools, resolves conflicts, and produces a ranked list of candidate locations. All tool invocations are constrained to data available prior to the target timestamp to ensure chronological validity. The entire process generates an auditable trace of reasoning steps and tool outputs.

Key Contributions

• Adaptive Evidence-Control Formulation: The paper recasts mobility prediction as an instance-level decision process where an agent chooses between a fast historical-regularity path and iterative evidence gathering for ambiguous cases.
• Training-Free Tool-Augmented Agent: AgentMob is a training-free framework that selectively invokes mobility tools (behavioral, geographical, and statistical) to produce reliable, auditable predictions without fine-tuning the underlying LLM.
• Strong and Explainable Performance: The framework demonstrates that adaptive evidence gathering significantly improves prediction quality, particularly when deterministic mobility statistics are insufficient or conflicting.

Experimental Results
The authors evaluated AgentMob on three datasets with varying spatial granularities: BW (Tokyo administrative polygons), YJMob100K (anonymized 500m grids), and Shanghai ISP (mobile network logs). They compared AgentMob against supervised baselines (DeepMove, Transformer) and several LLM-based methods (AgentMove, LLM-Mob, TrajLLM, LLM Urban Residents) using backbones like GPT-5.4, GPT-4.1-mini, and Qwen3-8B.

• Overall Performance: AgentMob achieved the strongest overall performance among all training-free LLM-based methods. With GPT-5.4, it reached 71.42% Acc@1 on BW, 33.14% on YJMob100K, and 33.50% on Shanghai ISP.
• Comparison with Supervised Models: While supervised models (like Transformer) remained superior on the BW dataset where regular data is abundant, AgentMob outperformed supervised baselines on YJMob100K and Shanghai ISP, suggesting that evidence-grounded reasoning is competitive when location spaces are dense or history is sparse.
• Role of the LLM Controller: An ablation study replacing the LLM controller with a deterministic rule (AgentMob-Statistics) showed that the LLM controller provides the most significant gains in ambiguous cases (non-fast-path). On the BW non-fast-path subset, the controller improved Acc@1 from 30.65% to 48.62%, demonstrating its value in cross-checking conflicting signals.
• Efficiency: AgentMob is highly efficient due to its adaptive strategy. On the BW dataset, 62.54% of samples were resolved via the fast path, resulting in lower wall-clock time and token usage compared to baselines that process all samples through a fixed, heavy inference procedure.

Significance and Claims
The paper claims that AgentMob represents a shift from static, one-pass prediction to adaptive, evidence-grounded, and iterative inference. Its primary significance lies in demonstrating that:

1. Adaptive Reasoning is Crucial: The ability to dynamically allocate reasoning effort based on prediction difficulty (fast path vs. deep reasoning) is more effective than forcing all samples through a uniform, computationally expensive procedure.
2. Evidence Grounding Improves Reliability: By explicitly retrieving and cross-checking multi-source evidence (behavioral, geographical, and statistical), the system can resolve ambiguities that static prompts or pure statistical models cannot.
3. Training-Free Viability: High-quality mobility prediction can be achieved without the cost of task-specific model training, provided the system is equipped with structured tools and an adaptive controller.

The authors conclude that the main benefit of their approach is not just raw accuracy, but the ability to provide reliable predictions in ambiguous scenarios where deterministic statistics fail, all while maintaining efficiency through a fast-path mechanism for routine cases.