Research map / Large language models / Retrieval and knowledge
Large Language Models as Urban Residents: An LLM Agent Framework for Personal Mobility Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LLMob, an LLM agent framework that generates individual daily activity trajectories by identifying self-consistent activity patterns and retrieving motivations, evaluated on Tokyo check-in data with GPT-3.5.
Problem
Personal mobility (activity trajectory) generation from real-world data is hard for data-driven models, which imitate data distributions without semantics and generalize poorly to unseen scenarios such as a pandemic.
Approach
An LLM agent with pattern identification (persona-conditioned candidate patterns selected by a self-consistency score against other residents' trajectories) and motivation retrieval (evolving-based via recent days, or learning-based via a contrastive date-similarity model) to condition daily trajectory generation.
Key results
- LLMob-E achieves the lowest DARD JSD in all three settings (e.g., 0.125 vs. best baseline around 0.318 in the normal setting), roughly 1/2 to 1/3 of the best baseline.
- LLMob-E attains the best step-interval JSD (0.046 in normal setting) while DeepMove is best on step distance (0.008 abnormal/0.013 pandemic).
- Removing pattern identification increases DARD JSD substantially in most settings, and self-consistency removal worsens DARD in most settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LLMob-E instead of the best baseline | personal mobility generation · Normal trajectory, normal data | SI: 0.058 -> 0.046 | Helped |
| Use LLMob-E instead of the best baseline | personal mobility generation · Normal trajectory, normal data | DARD: 0.318 -> 0.125 | Helped |
| Use LLMob-E instead of the best baseline | personal mobility generation · Abnormal trajectory, abnormal data | SI: 0.060 -> 0.043 | Helped |
| Use LLMob-E instead of the best baseline | personal mobility generation · Abnormal trajectory, abnormal data | DARD: 0.339 -> 0.127 | Helped |
| Use LLMob-E instead of the best baseline | personal mobility generation · Abnormal trajectory, normal data | SI: 0.058 -> 0.056 | Helped |
| Use LLMob-E instead of the best baseline | personal mobility generation · Abnormal trajectory, normal data | DARD: 0.218 -> 0.117 | Helped |
| Add activity patterns (ablation) | LLMob-E · Normal trajectory, normal data | DARD: 0.223 -> 0.125 | Helped |
| Add activity patterns (ablation) | LLMob-E · Abnormal trajectory, abnormal data | DARD: 0.252 -> 0.127 | Helped |
| Add activity patterns (ablation) | LLMob-E · Abnormal trajectory, normal data | DARD: 0.209 -> 0.117 | Helped |
| Add self-consistency evaluation (ablation) | LLMob-E · Normal trajectory, normal data | DARD: 0.295 -> 0.125 | Helped |
| Add self-consistency evaluation (ablation) | LLMob-E · Abnormal trajectory, abnormal data | DARD: 0.225 -> 0.127 | Helped |
| Add self-consistency evaluation (ablation) | LLMob-E · Abnormal trajectory, normal data | DARD: 0.301 -> 0.117 | Helped |
Surprising findings
- Without self-consistency (random pattern selection), performance is sometimes worse than removing patterns entirely.
- LLMob does not excel at spatial step distance; DeepMove and TrajGAIL win on spatial metrics.
Limitations
- Only a single dataset (Tokyo, 100 users) was used for the main evaluation.
- Only GPT-3.5 was fully evaluated due to cost; individual agents modeled without interactions.
- LLMob is weaker on spatial step-distance metrics than some baselines.
Experimental setup
Datasets: Tokyo personal activity trajectory dataset (Twitter and Foursquare check-ins, 2019-2022)
Benchmarks: Tokyo personal activity trajectory dataset (normal 2019, abnormal 2020, pandemic 2021)
Metrics: Jensen-Shannon divergence on step distance step interval daily activity routine distribution spatial-temporal visits distribution
Baselines: Markov-based mechanic model (MM) LSTM DeepMove STAN TrajGAIL ActSTD DiffTraj
Builds on: DeepMove TrajGAIL DiffTraj InfoNCE
Closest papers by meaning
- Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.85)
- FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents (ICLR 2026 · similarity 0.85)
- TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration (NeurIPS 2025 · similarity 0.85)
- RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility (NeurIPS 2025 · similarity 0.85)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.84)
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.