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

Source: arXiv - 2606.05130v1

Overview

Predicting where an individual will go next is a cornerstone for smart‑city applications, from traffic‑flow optimization to personalized services. The paper Towards Efficient and Evidence‑grounded Mobility Prediction with LLM‑Driven Agent introduces AgentMob, a novel framework that lets a large language model (LLM) act like an autonomous “investigator” to decide the next location without any task‑specific training. By combining fast, rule‑based shortcuts with on‑the‑fly evidence gathering, AgentMob pushes the accuracy of training‑free LLM approaches to new heights while keeping the decision process transparent.

Key Contributions

• Training‑free LLM‑driven agent that treats mobility prediction as an adaptive decision‑making problem rather than a static generation task.
• Hybrid fast‑path / evidence‑controlled workflow: routine predictions are resolved instantly using historical regularity; ambiguous cases trigger iterative tool usage (e.g., querying recent trajectories, stay‑move likelihood, geographic context).
• Evidence‑grounded reasoning: the LLM explicitly selects and combines multiple sources of mobility evidence, making its predictions interpretable to developers and analysts.
• State‑of‑the‑art performance among training‑free methods: with GPT‑5.4, AgentMob reaches 71.42 % top‑1 accuracy on the BW dataset and competitive scores on YJMob100K and Shanghai ISP.
• Open‑source implementation (GitHub) that can be plugged into existing pipelines.

Methodology

1. Problem Framing – The next‑location task is cast as a sequential decision problem: given a user’s recent trajectory, decide whether the next move is “routine” (highly predictable) or “ambiguous”.
2. Fast‑Path Module – A lightweight statistical check (e.g., frequency of the last visited POI, time‑of‑day regularity) instantly returns a prediction for routine cases, saving LLM inference time.
3. LLM Controller – When the fast‑path confidence falls below a threshold, the LLM (e.g., GPT‑5.4) becomes the controller. It can invoke a toolbox of evidence tools, each exposing a specific data view:
   • Recent Trajectory Tool: returns the last N visited locations.
   • Historical Behavior Tool: aggregates long‑term visitation patterns for the user.
   • Stay‑Move Likelihood Tool: computes the probability of staying vs. moving given context (time, day).
   • Geographical Evidence Tool: provides POI categories, distance metrics, or traffic conditions for candidate locations.
4. Iterative Evidence Gathering – The LLM decides which tool(s) to call, inspects the returned evidence, and may request additional data in a loop until it reaches a confidence threshold.
5. Decision Output – Finally, the LLM produces a ranked list of candidate locations, with the top‑1 entry used as the prediction.

The whole pipeline requires no fine‑tuning of the LLM; all adaptation happens at inference time through prompt engineering and tool selection.

Results & Findings

• Overall superiority: AgentMob outperforms all other training‑free LLM baselines across the three benchmarks.
• Evidence advantage: In the BW dataset, the LLM controller lifts accuracy on hard cases from 30.65 % (a statistical baseline using the same tools) to 48.62 %, confirming that adaptive evidence selection is the main driver of improvement.
• Efficiency: The fast‑path handles the majority of routine predictions, keeping average latency low while still allowing the LLM to focus on the truly ambiguous instances.

Practical Implications

• Plug‑and‑play mobility services – Companies can integrate AgentMob into ride‑hailing, bike‑share, or logistics platforms without collecting a massive labeled training set.
• Explainable predictions – Because the LLM logs which evidence tools were consulted, developers can surface “why” a location was chosen (e.g., “high stay probability at nearby office during lunch hour”). This aids compliance with emerging AI‑transparency regulations.
• Resource‑aware deployment – The fast‑path reduces API calls to expensive LLM endpoints, lowering operational costs while preserving high accuracy where it matters most.
• Cross‑domain adaptability – The same framework can be repurposed for other sequential decision tasks (e.g., next‑item recommendation, anomaly detection) by swapping in domain‑specific evidence tools.
• Rapid prototyping – Since no fine‑tuning is required, data scientists can experiment with new evidence sources (weather, events, public‑transport schedules) simply by adding a new tool and updating prompts.

Limitations & Future Work

• Dependence on LLM capabilities – The approach’s ceiling is tied to the underlying model’s reasoning power; older or smaller LLMs may not yield the same gains.
• Tool design overhead – Crafting high‑quality evidence tools (especially for heterogeneous city data) still requires engineering effort and domain knowledge.
• Scalability of iterative loops – In worst‑case scenarios, the LLM may request many tools, increasing latency; smarter stopping criteria are an open research direction.
• Evaluation on privacy‑sensitive data – The current experiments use public datasets; real‑world deployment must address user privacy and data‑governance constraints.

Future work suggested by the authors includes:

1. Learning a meta‑controller that automatically balances fast‑path vs. LLM control.
2. Extending the toolbox with real‑time traffic and event streams.
3. Exploring few‑shot fine‑tuning to further boost performance while retaining interpretability.

Authors

• Linyao Chen
• Qinlao Zhao
• Zechen Li
• Mingming Li
• Likun Ni
• Jinyu Chen
• Yuhao Yao
• Xuan Song
• Noboru Koshizuka
• Hiroki Kobayashi

Paper Information

• arXiv ID: 2606.05130v1
• Categories: cs.LG, cs.AI
• Published: June 3, 2026
• PDF: Download PDF