CityReal: Human-Aligned Urban Behavior and City Dynamics Simulation with Large-Scale LLM Agents

Nicolas Bougie; Xiaotong Ye; Narimasa Watanabe

Source-linked AI summary

CityReal: Human-Aligned Urban Behavior and City Dynamics Simulation with Large-Scale LLM Agents

Nicolas Bougie, Xiaotong Ye, Narimasa Watanabe

arXiv:2608.16897v1Jul 8, 2026physics.soc-phcs.AIcs.MA

TL;DR

Existing LLM-based urban simulators often rely on few-shot prompts without aligning decisions to observed human behavior or maintaining coherent, experience-based routines. CityReal uses intention-driven agents, experience-based adaptation, and learned textual adapters, improving alignment with real-world behavior at both individual and population levels while supporting large-scale urban analysis.

Problem
Existing LLM-based urban simulators often lack alignment with observed human behavior, continuity across decisions, and learning from agents’ own histories.
Method
CityReal combines intention-driven planning, experience-based adaptation, and learned textual adapters that calibrate behavior modules to target population statistics.
Results
CityReal agents closely align with human behavior at both micro and macro levels and support more realistic urban phenomena modeling than prior agent-based models.
Takeaways & Limitations
CityReal provides a scalable foundation for studying complex urban phenomena across behavioral modeling, urban planning, and social studies.
Takeaways & Limitations
Some experiments use non-public datasets, limiting direct comparison and independent verification.

Abstract

from arXiv · show

Large-scale urban simulation plays a pivotal role in social science, traffic safety, and transportation policy. Recent work has shown that large language models, when prompted as agents, can generate lifelike daily routines at city scale. Yet these methods typically rely on few-shot prompting, causing agents to reproduce the LLM's behavioral priors rather than the target population. We introduce CityReal, a modular framework for human-aligned urban simulation. CityReal models agents as intention-driven decision makers that pursue coherent mobility and activity plans rather than isolated step-by-step choices. They adapt over time by learning habits and preferences based on experience and constraints. To improve population-level realism, we learn textual adapters for behavior modules that align agent decisions with observed population statistics. Experiments show that CityReal improves alignment with real-world human behavior at both micro and macro levels. Scaling to tens of thousands of agents, it supports analysis of crowd density, place popularity, mobility flows, and well-being under different urban scenarios, offering a scalable testbed for urban simulation and forecasting.

1 Introduction

CityReal addresses limitations in LLM-based urban simulation by aligning agent decisions with observed human patterns while organizing behavior around persistent intentions and experience-driven adaptation. Its modular design targets coherent individual behavior and population-level realism without requiring updates to the underlying LLM.

Motivation: Traditional urban simulators use transparent, scalable hand-crafted rules but produce rigid behavior that adapts poorly to unseen situations.These limitations motivate more flexible approaches to reproducing daily urban behavior.
Related work: LLM-based urban agents reason in natural language and condition actions on demographic profiles, internal states, and surrounding environments.Recent urban simulation systems use this paradigm to generate resident trajectories and study collective social phenomena.
Problem: Existing LLM-based urban simulators often use few-shot prompting without explicitly aligning agent decisions with observed human behavior.They also make discretionary decisions as isolated choices, yielding trajectories with limited continuity compared with behavior organized around persistent intentions.
CityReal: CityReal learns textual adapters with Monte Carlo Tree Search to calibrate agent decisions toward target population statistics while keeping the LLM frozen.The framework combines this alignment mechanism with intention-driven behavior and experience-based adaptation.
CityReal: CityReal maintains stable motives across consecutive actions and uses end-of-day reflection to accumulate habits, preferences, social tendencies, and constraints.These mechanisms operate with persona, memory, belief, mobility, and social modules to support adaptive urban agents.

2 Related Work

Related work spans traditional agent-based models and newer LLM-based urban simulations, but existing approaches face limitations in behavioral realism, population alignment, and computational efficiency. Recent methods add memory, place-selection, recursive planning, and belief modules while scaling urban simulation.

Traditional Agent-Based Models: Traditional agent-based models study complex phenomena, resource allocation, and policy evaluation but rely on hard-coded rules or fixed utility functions.These design choices constrain their ability to capture behavior.
Scalable Urban Simulation: Computational efficiency becomes increasingly important as urban simulations scale to larger populations.Park et al. (2024) scale simulations to 1,000 agents but still incur prohibitive costs.
LLM-Based Urban Simulation: Recent LLM-based approaches improve simulation realism using episodic memory, gravity-based place selection, recursive activity planning, dynamic memory, and belief modules.AgentSociety, MobileCity, and CitySim represent different combinations of these mechanisms, yet existing approaches still struggle to align simulated populations with observed human behavior.

3 Method

CityReal simulates urban residents as coherent, intention-driven agents embedded in a graph-structured city, using personas, memories, beliefs, schedules, mobility, and social interactions. Textual adapters calibrate behavior modules toward target population statistics without modifying the per-agent reasoning loop or retraining the underlying LLM.

Framework overview: CityReal embeds persona-initialized agents in a graph-structured city and learns textual adapters that condition behavior modules toward target population statistics.City nodes represent places or areas, while edges denote mobility connections such as car or train routes.
Agent representation: Personas encode demographic attributes, spatial anchors, and psychographic traits that constrain feasible activities and daily routines.Examples include school schedules for students, weekday commuting for full-time workers, and flexible routines for retirees.
Agent representation: Memory and belief modules record experiences and update subjective evaluations of places and social contacts, shaping later destination and interaction decisions.Place evaluations include affordability, convenience, crowding, and enjoyment; social beliefs include affinity, trust, and familiarity.
Activity planning: Daily planning assigns mandatory and medium-priority tasks recursively, while flexible blocks are instantiated at execution time through natural-language intentions.Each intention describes what the agent is trying to accomplish and may include an area anchor and expected duration.
Population alignment: Calibration steers decisions toward population-level distributions with module-specific textual adapters, preserving the reasoning loop while penalizing implausibility and reduced diversity.Adapters are optimized iteratively using improvement signals, then transferred from representative persona types to uncalibrated agents by nearest-neighbor persona matching.

4 Experiments

Experiments evaluate CityReal against established and recent agent-based baselines using 3,000 agents in Tokyo. Results indicate alignment with population activity patterns, human-like routines, travel distributions, POI popularity, well-being estimates, and crowd-density patterns.

Settings and baselines: Experiments use 3,000 agents in the Tokyo metropolitan area and compare CityReal with GeAn, AGA, HumanoidAgent, MobileCity, AgentSociety, and CitySim.Agents are powered by GPT-5.4-mini unless otherwise specified.
Population activity patterns: Across two simulated months, CityReal closely matches 2021 Japanese survey distributions of daily time spent across age groups and activity categories.The categories are Work, Commute, Housework, and Personal Care & Sleep.
Individual behavior: Pairwise routine judgments evaluate Naturalness, Coherence, and Plausibility after stylistic normalization, while reflective memory and intention modeling yield more socioeconomically consistent routines.The evaluation uses 15 independent trials per approach and GPT-5 judgments.
Travel distributions: CityReal closely reproduces weekday and weekend travel distributions by calibrating behavior to observed data and using mandatory schedule planning to produce consistent commute peaks.Prior methods primarily generate travel decisions from prompt-driven agent decisions.
POI popularity: CityReal achieves stronger correlations between simulated and real-world POI popularity in Shibuya, using agent visits over one simulated month against Google Maps-based ground truth.The evaluation reports Spearman rank correlations.
Well-being estimation: CityReal outperforms all agent-based baselines for five-class well-being prediction, although XGBoost achieves the highest macro F1-score.The evaluation uses 1,200 Japanese survey responses and compares simulated questionnaire answers with an XGBoost baseline and prior agent-based methods.
Crowd density: CityReal accurately reproduces Shibuya crowd-density patterns, including the highest concentrations around the train station and major commercial streets.Simulated heatmaps aggregate agent visits and are compared with distributions estimated from smartphone location data; the intention module helps keep agents within a target area.

5 Conclusion

CityReal is a modular framework for LLM-powered simulation of human-aligned urban behavior. It combines coherent intention-driven decisions, experience-based adaptation, and learned calibration to support realistic urban research and applications.

CityReal is a modular framework for simulating human-aligned urban behavior with LLM-powered agents.
CityReal targets individual-level coherence and population-level alignment by maintaining intentions, learning from experience, and calibrating behavior modules with learned adapters.The adapters align behavior-generating modules with real-world statistics.
CityReal supports realistic agent behaviors and the study of complex urban phenomena across behavioral modeling, urban planning, and social studies.The conclusion presents it as a foundation for research and industry applications.

6 Limitations

CityReal’s limitations include constrained reproducibility from non-public datasets and reliance on large l…. Ablation studies partially address this issue, but finer-grained analysis of module interactions remains future work.

Limitations: Non-public datasets constrain reproducibility, limiting direct comparison and independent verification.Because most evaluations use macro-level aggregate statistics, future methods can still compare against CityReal using the reported targets.
Limitations: Reliance on large l… remains a limitation that the authors only partially address through ablation studies.
Limitations: More fine-grained analysis of module interactions remains an important direction for future work.

7 Ethics Statement

The paper identifies ethical risks in large-scale synthetic urban-agent simulations, including demographic bias and behavior-steering uses. It recommends transparency, bias evaluation, human oversight, and using simulations to complement rather than replace human expertise.

Ethical risks: Synthetic agents may reproduce or amplify biases from the underlying LLM, initialization process, or calibration data across demographic and lifestyle factors.These biases could overrepresent some groups while underrepresenting or disadvantaging others in urban-planning or policy analyses.
Ethical risks: Large-scale behavioral simulations could identify patterns that support interventions steering collective behavior, raising concerns without transparency and oversight.The passage specifically flags potential concerns about such uses, though the excerpt does not provide further detail.
Responsible use: Synthetic agents support early-stage exploration and low-cost urban-scenario evaluation but should not substitute for residents, stakeholders, or domain experts.Simulation abstracts away lived experience, local knowledge, and contextual factors that may not be captured by an LLM-based system.
Responsible use: The authors recommend developing and deploying LLM-based urban simulations with transparency, bias evaluation, human oversight, and attention to social consequences.They frame synthetic agents as complements to human participation rather than replacements.

A Experimental Setup … B Population-Level Alignment

CityReal initializes diverse agents from survey-aligned personas, simulates intention-driven daily behavior with memory, needs, social interaction, and reflection, and calibrates modules toward population statistics using textual adapters. The experimental setup specifies reproducible simulation settings and evaluation criteria, while population-level alignment rewards reduced aggregate discrepancy without sacrificing plausibility or diversity.

A Experimental Setup: Agents receive survey-initialized personas covering demographics, household and life-stage information, occupations, finances, locations, and psychographic traits.The attribute distributions closely match recent Japanese census statistics and lifestyle surveys.
A Experimental Setup: CityReal tracks temporal and spatial memories, social beliefs, short-term needs, and continuous financial pressure to condition behavioral decisions.Temporal memory retrieves the top k1 = 5 entries from the past ∆t = 24 hours, while unvisited-place beliefs use k = 10 similar visited locations.
A Experimental Setup: The simulation uses 5-minute timesteps, fixed random seeds, and daily schedule blocks with a minimum 5-minute granularity.Frontier LLM APIs may still exhibit minor backend-level nondeterminism.
A.1 Module Details: CityReal comprises LLM-based agents with diverse personas that plan daily activities, interact socially, and navigate a virtual city environment.The overview figure summarizes the framework’s agent-environment organization.
A.1.1 Planning and Intention Module: Daily planning recursively assigns fixed activities and fills remaining intervals with routines, while active intentions preserve shared context across subsequent decisions.The orchestrator can continue, revise, complete, or replace an intention when activities change, needs become urgent, or the environment changes.
A.1.1 Planning and Intention Module: Place and transport selection combine candidate filtering with memory-based beliefs and contextual constraints, while face-to-face interaction is limited to one partner per 30-minute window.The area selector considers the top 10 candidate areas, the POI selector up to 200 POIs, and the gravity model uses γ = 2.0.
A.1.2 Reflection Module: At each simulated day’s end, the reflection module processes the trajectory and retrieved memories to generate up to four structured candidate reflections.Inputs include up to 24 representative temporal memories and the top-8 relevant existing reflective memories.
B Population-Level Alignment: Population-level alignment searches textual adapters for behavior modules by comparing aggregate simulation statistics with targets, while penalizing plausibility loss and diversity collapse.The initial adapter state Φ0 is empty, lower R(Φt) indicates better alignment, and positive return means reduced discrepancy.

B.1 Alignment Objective Details · B.2 Adapter Edit Vocabulary · B.3 Search Procedure

CityReal aligns population-level behavior using measure-specific discrepancies and penalties for implausible interruptions or insufficient behavioral diversity. It searches constrained textual adapter edits with MCTS, then transfers adapters found on a representative 500-agent subset to the full population.

B.1 Alignment Objective Details: Continuous measures use log-space 1-Wasserstein distance with weighted CCDF L1, categorical measures use Jensen–Shannon divergence, and scalar summaries use average relative error.Examples include travel distance, inter-trip interval, transport distribution, origin–destination flows, trip purpose, and wake-up time.
B.1 Alignment Objective Details: The plausibility penalty tracks emergency breaks caused by critical sleep, food, or health needs and raises the objective when alignment produces excessive interruptions.The threshold is set relative to the unadapted baseline as ρthr = max(1.5, ρ0 + 1.0), with λp = 0.5.
B.3 Search Procedure: MCTS selects nodes with UCT, evaluates expanded candidates through rollouts, back-propagates cumulative returns, and retains all evaluated candidates as revisitable siblings.Global statistics for each (module, edit) pair help effective edits generalize across agent groups.
B.1 Alignment Objective Details: The diversity penalty groups agents by bucketed persona and behavior features, requires at least five agents per group, and evaluates decision entropy by group and module.Features include age_bin, household, occupation, trips_per_day, and local_trip_rate.
B.2 Adapter Edit Vocabulary: Each search action selects one target module and up to two axis–direction pairs, which a rewriter converts into a textual adapter; every axis also has a neutral option.Only the allowed edit axes are retained during candidate generation.
B.3 Search Procedure: The analyst LLM proposes module- and group-specific edits by comparing simulated with target statistics, retaining only allowed edits covering sufficiently large groups.Candidate generation is constrained by the edit vocabulary in Table 2.
B.3 Search Procedure: The search uses UCT constant 2, global exploration constant 1.0, candidate pool size 3, maximum depth 4, and at most 20 iterations.It stops when the best rollout reward improves by no more than 0.005 over five consecutive iterations; temperatures are 0.1 for analyst/rewriter LLMs and 0.2 for rollouts.
B.3 Search Procedure: Adapters are optimized on 500 agents covering major demographic and spatial types, then transferred by matching each remaining agent to the closest subset agent in personaembedding space.The subset spans age, household, occupation, and spatialanchor types.

C Discussion … E.4 Adapter Convergence and Efficiency

The discussion identifies limitations in CityReal’s interpretability, data dependence, bias, and evaluation, while describing a hybrid design for transparent simulation. Experiments and the simulation loop show how CityReal calibrates, plans, adapts, and evaluates human-aligned urban behavior.

C Discussion: CityReal’s individual decision reasoning remains partially opaque because large language models are black boxes, despite producing collective behaviors consistent with urban mobility patterns.The paper notes that training corpora may encode urban routines and cultural patterns, making learned norms difficult to disentangle from genuine generalization.
C Discussion: Behavioral grounding depends on socioeconomic and local data quality, so coarse or sparse records can bias financial constraints, personas, spatial anchors, and downstream conclusions.Salary-conditioned personas, spending tracking, and financial-pressure estimates improve grounding but remain approximations when income, expenditure, demographic, spatial, or economic data are limited.
C Discussion: Population-level alignment reduces some demographic and cultural bias, but aggregate calibration can hide minority differences, reproduce flawed targets, and downweight rare realistic behaviors.The paper also notes that LLM-as-judge evaluation may contain stylistic bias and remains imperfect despite cross-model validation with Gemini-3.5.
C Discussion: CityReal combines contextual LLM decisions with explicit models for financial pressure, beliefs, needs, and other latent states to improve efficiency, interpretability, and auditability.Explicit models were easier to inspect and control than LLMs for scalar internal states, while needs follow established psychological frameworks and financial pressure uses tracked income and expenditure.
D Pseudo-Code: The simulation loop calibrates adapters, initializes agents, plans daily blocks, updates intentions and states, executes activities and movement, reflects on experience, and aggregates urban outcomes.Inputs are the city environment, agents, and target population statistics; outputs are trajectories and aggregate outcomes involving mobility, crowds, popularity, and well-being.
E.1 Human Likeliness: 20,000 outputs per method across four domains are scored on a 5-point Likert scale, and CITYREAL achieves the highest human-likeness scores across all domains.The largest gains occur for mobility and event reactions, while dialogue gains are smaller; the unaligned variant also outperforms CITYSIM.
E.2 Belief Estimation: Larger models achieve lower belief-estimation errors overall when predicting beliefs for unvisited POIs across five semantic categories.Accuracy is measured by mean absolute error between predicted and ground-truth beliefs, and these beliefs condition destination choice and activity planning.
E.3 Population-level Time-use Alignment: CITYREAL reduces Jensen–Shannon divergence versus CITYSIM across all age groups while preserving age-specific behavioral patterns rather than collapsing them into one population distribution.The comparison uses simulated time-use distributions against the 2021 Japanese national time-use survey.

E.5 Robustness to Evaluator Bias … E.8 Effect of Search Subset Size

CITYREAL remains preferred when evaluated by Gemini-3.5, supports counterfactual weather analysis in Shibuya, and is assessed through component ablations and search-subset experiments. Larger alignment-search subsets improve population-level calibration, with diminishing returns after major persona types are covered.

E.5 Robustness to Evaluator Bias: CITYREAL’s overall ranking remains stable when Gemini-3.5, from a different model family, evaluates pairwise human-likeness.This tests whether using GPT as both evaluator and agent generator introduces model-family bias.
E.6 Urban Policy A/B Testing: Rain shifts Shibuya activity from outdoor leisure and walking toward indoor, stay-home, and station-adjacent areas.The comparison uses the same agents and environment under cloudy and rainy conditions.
E.6 Urban Policy A/B Testing: Weather-only counterfactuals enable analysis of demand, crowding, movement patterns, event plans, station congestion, retail foot traffic, and infrastructure stress.Because scenarios differ only in weather, outcome changes reflect agents’ adaptation to environmental conditions.
E.7 Ablation Study: The ablation study removes population-level alignment, financial pressure, social interaction, reflection, intention formation, recursive planning, beliefs, needs, and persona one component at a time.It evaluates individual realism using human-likeness across four domains and population alignment using Jensen–Shannon divergence.
E.8 Effect of Search Subset Size: Alignment gains diminish once the main persona types are covered, so full-population search provides limited additional benefit and CITYREAL uses 500 agents.The search learns adapters on a subset and transfers them to the remaining population.
E.7 Ablation Study: Ablation evaluation measures human-likeness across activity, dialogue, mobility, and event reaction, alongside Jensen–Shannon divergence for calendar and vehicle distributions.Human-likeness uses GPT-5 Likert ratings from 1–5, while lower divergence indicates better population-level alignment.
E.8 Effect of Search Subset Size: Increasing the alignment-search subset size consistently improves population-level alignment, while small subsets underrepresent minority persona groups.The resulting calibration weakness is especially pronounced for transport-mode distribution.

F Cost Analysis … I Simulation Interface

The paper reports CityReal’s one-time calibration and recurring simulation costs, compares its agent, spatial, and mobility capabilities with prior systems, evaluates human-likeness using an LLM judge, and provides an interface for inspecting city dynamics.

F Cost Analysis: CityReal’s default adapter calibration costs approximately $248 once for a target population and can be reused across agents, days, and scenarios.Calibration uses a subset of agents with short simulation rollouts.
G Comparison with Prior Work: The comparison with prior urban simulation systems covers alignment, adaptive memory, intention modeling, financial state, social interaction, scalability, spatial reasoning, and mobility reasoning.Tables 7 and 8 summarize agent-level capabilities and spatial and mobility reasoning, respectively.
H LLM Evaluator Prompt: An LLM-based evaluator judges whether simulated urban behavior traces resemble AI-generated or real-human behavior on a 1-to-5 scale.The evaluator receives interaction logs and assigns 1 as most like AI and 5 as most like a human.
H LLM Evaluator Prompt: CityReal is the only compared framework jointly supporting population-level alignment, persistent intentions, financial state modeling, and adaptive memory at large scale.This capability comparison is reported in Table 7.
H LLM Evaluator Prompt: CityReal extends prior mobility models with intention-aware area selection, belief-aware gravity scoring, transport choice, and cost-aware decision making.These spatial and mobility reasoning capabilities are summarized in Table 8.
I Simulation Interface: The interactive simulation interface visualizes trajectories, activities, intentions, schedules, mobility patterns, and context in real time.Users can inspect individual agents, compare populations, and analyze routine evolution under environmental conditions or interventions.