Urban-MAS: LLM-Based Urban Prediction Framework

• Urban-MAS is an LLM-based multi-agent system framework designed to achieve zero-shot human-centered urban predictions through coordinated agent orchestration.
• It employs a three-layer architecture—Predictive Factor Guidance, UrbanInfo Extraction, and Multi-UrbanInfo Inference—to systematically prioritize and integrate diverse urban data.
• Empirical results from Tokyo, Milan, and Seattle demonstrate notable error reductions in urban perception tasks, validating the framework’s structured inference approach.

Searching arXiv for the specified Urban-MAS paper and closely related urban multi-agent systems work. Urban-MAS is an LLM-based Multi-Agent System framework for human-centered urban prediction under zero-shot settings. It is designed to improve zero-shot human-centered urban predictions by prioritizing the most predictive factors, extracting reliable urban information, and integrating multi-source signals into a final inference. The framework is situated within Urban Artificial Intelligence, where LLMs can integrate multimodal inputs to address heterogeneous data in complex urban systems but often underperform on domain-specific tasks. In experiments on running-amount prediction and urban perception across Tokyo, Milan, and Seattle, Urban-MAS substantially reduces errors relative to a GPT-5 single-LLM baseline (Lou, 30 Oct 2025).

1. Problem setting and conceptual scope

Urban-MAS addresses human-centered urban tasks such as perception prediction and human dynamics by targeting a specific difficulty in LLM-based urban inference: heterogeneous urban evidence can be ingested by a general-purpose model, yet domain-specific predictive performance may remain limited. The framework therefore treats prediction not as a single prompt-response operation, but as a coordinated multi-agent process in which factor prioritization, information extraction, and inference are explicitly separated (Lou, 30 Oct 2025).

The prediction task is denoted by $\tau$, with examples including “estimate running amount” and “predict perceived liveliness.” For a single location $\ell$, the framework constructs a sequence of intermediate representations rather than directly emitting a final score. This design makes Urban-MAS a structured zero-shot prediction architecture rather than a conventional fine-tuning pipeline. A plausible implication is that the framework is intended to compensate for weak domain specialization in foundation models by imposing agent-level decomposition over the urban reasoning process.

2. Three-layer agent architecture

Urban-MAS is a three-layered multi-agent system. Its first layer consists of Predictive Factor Guidance Agents, its second of Reliable UrbanInfo Extraction Agents, and its third of Multi-UrbanInfo Inference Agents. These three layers correspond to factor prioritization, reliability enhancement, and multi-source inference (Lou, 30 Oct 2025).

In the first layer, Urban-MAS allocates Deep-research subagents to each pair of dimension $d \in \{\text{Social}, \text{Built Environment}\}$ and scale $r \in \{\text{Macro}, \text{Street}\}$. Each Deep-research subagent produces a short report $t_{(d,r)}$ listing six candidate factors. A Summary subagent then compresses $t_{(d,r)}$ into an ordered set of top-$k$ predictive factors,

$P_{(d,r)} = \{f_1, f_2, \ldots, f_k\},$

usually with $k=6$. The paper does not provide a closed-form feature-importance formula, but writes symbolically

$$\mathrm{score}(f) = \mathrm{RankWeight}(f; t_{(d,r)}),$$

where $\ell$0 assigns descending scores so that $\ell$1 has the highest weight. The union over dimensions and scales serves as a targeted guidance mask that steers subsequent knowledge extraction toward the most influential factors for task $\ell$2.

In the second layer, an Extractor subagent produces two JSON-structured variants of UrbanInfo, $\ell$3 and $\ell$4, each keyed by the fields in $\ell$5. An Evaluator subagent compares these outputs at the field level, and a Refiner subagent re-extracts only those fields where consistency is insufficient. This layer is intended to improve robustness by comparing multiple outputs, validating consistency, and re-extracting when conflicts occur.

In the third layer, reliable outputs from all four dimension-scale combinations are concatenated under a strict schema and passed to a single Inference Agent, which prompts GPT-5 in JSON mode to produce the final prediction $\ell$6. This architecture makes the final prediction contingent on a prior process of structured evidence collection rather than on a monolithic prompt.

3. Algorithmic pipeline and mathematical formulation

For each location $\ell$7 and task $\ell$8, the pipeline begins with factor discovery. For every $\ell$9 pair, a Deep-research Agent generates a report on top-6 factors, and a Summary Agent converts that report into a predictive factor set $d \in \{\text{Social}, \text{Built Environment}\}$0. The next stage performs extraction conditioned on those factors. For each $d \in \{\text{Social}, \text{Built Environment}\}$1, the Extractor produces two variants, the Evaluator computes similarity field-wise, and the Refiner is invoked only when similarity falls below threshold. The final stage integrates the conflict-free outputs into a single prediction (Lou, 30 Oct 2025).

The field-level consistency criterion is a “hybrid soft similarity”:

$d \in \{\text{Social}, \text{Built Environment}\}$2

Here,

$d \in \{\text{Social}, \text{Built Environment}\}$3

and $d \in \{\text{Social}, \text{Built Environment}\}$4 measures longest matching substrings. If $d \in \{\text{Social}, \text{Built Environment}\}$5 with $d \in \{\text{Social}, \text{Built Environment}\}$6, $d \in \{\text{Social}, \text{Built Environment}\}$7 is accepted as $d \in \{\text{Social}, \text{Built Environment}\}$8. Otherwise, the Refiner re-extracts only those fields where $d \in \{\text{Social}, \text{Built Environment}\}$9, producing conflict-free, reliable output $r \in \{\text{Macro}, \text{Street}\}$0.

Once reliable JSON outputs are available,

$r \in \{\text{Macro}, \text{Street}\}$1

the Inference Agent concatenates them under a strict schema, exemplified by {"running_amount": float}, and prompts GPT-5 in JSON mode to produce the final prediction. The joint inference is written symbolically as

$r \in \{\text{Macro}, \text{Street}\}$2

where $r \in \{\text{Macro}, \text{Street}\}$3 denotes the LLM’s internal reasoning.

This decomposition has an important methodological consequence. The framework does not attempt to make the LLM internally more specialized; instead, it constrains the LLM’s operating context through ranked factors, redundancy-based extraction, and schema-controlled inference. This suggests a form of externalized control over zero-shot urban reasoning.

4. Evaluation protocol and empirical performance

Urban-MAS was evaluated on 300 samples drawn from Tokyo, Milan, and Seattle for two tasks: running-amount prediction, defined as Strava heatmap brightness scaled to $r \in \{\text{Macro}, \text{Street}\}$4, and urban perception, defined through PlacePulse liveliness and boringness scores in $r \in \{\text{Macro}, \text{Street}\}$5. GPT-5 used in zero-shot mode serves as the single-LLM baseline. Performance is measured with Mean Absolute Error, Mean Squared Error, and Root Mean Squared Error:

$r \in \{\text{Macro}, \text{Street}\}$6

$r \in \{\text{Macro}, \text{Street}\}$7

$r \in \{\text{Macro}, \text{Street}\}$8

Lower values are better (Lou, 30 Oct 2025).

Task / Metric	Single-LLM	Urban-MAS
Running Amount / MAE	2.99	2.97
Running Amount / MSE	13.73	13.20
Running Amount / RMSE	3.70	3.63
Boringness / MAE	2.83	2.05
Boringness / MSE	9.95	5.84
Boringness / RMSE	3.15	2.42
Liveliness / MAE	2.69	1.73
Liveliness / MSE	9.10	4.40
Liveliness / RMSE	3.02	2.10

The corresponding relative reductions are reported as follows: for Running Amount, MAE $r \in \{\text{Macro}, \text{Street}\}$9 $t_{(d,r)}$0, MSE $t_{(d,r)}$1 $t_{(d,r)}$2, and RMSE $t_{(d,r)}$3 $t_{(d,r)}$4; for Urban Perception (Boringness), MAE $t_{(d,r)}$5 $t_{(d,r)}$6, MSE $t_{(d,r)}$7 $t_{(d,r)}$8, and RMSE $t_{(d,r)}$9 $t_{(d,r)}$0; for Urban Perception (Liveliness), MAE $t_{(d,r)}$1 $t_{(d,r)}$2, MSE $t_{(d,r)}$3 $t_{(d,r)}$4, and RMSE $t_{(d,r)}$5 $t_{(d,r)}$6. The reported pattern is that Urban-MAS consistently outperforms the single-LLM baseline across all metrics, with particularly large gains on perception tasks.

These results delimit the framework’s strongest empirical claim. Error reduction is modest for running-amount prediction but pronounced for boringness and liveliness. A plausible implication is that the structured extraction and multi-source integration pipeline is especially beneficial when the target variable is more weakly specified and more dependent on heterogeneous urban cues.

5. Ablation evidence and component attribution

Two ablations were performed to quantify the contribution of each agent layer. The first, –PredictiveFactors, disables the Predictive Factor Guidance layer by using generic prompts. The second, –ReliabilityBoost, disables the two-variant consistency check so that only a single Extractor output is used. Relative error increases are measured against full Urban-MAS (Lou, 30 Oct 2025).

Task / Ablation	MAE $t_{(d,r)}$7	MSE $t_{(d,r)}$8
Running Amount / –PredictiveFactors	+52.84%	+102.98%
Running Amount / –ReliabilityBoost	+0.30%	+1.46%
Boringness / –PredictiveFactors	+16.37%	+28.69%
Boringness / –ReliabilityBoost	+11.52%	+19.45%
Liveliness / –PredictiveFactors	+46.89%	+83.79%
Liveliness / –ReliabilityBoost	+28.06%	+47.02%

The corresponding RMSE increases are +42.47% and +0.73% for Running Amount, +13.44% and +9.29% for Boringness, and +35.57% and +21.25% for Liveliness, respectively. The paper’s key observations are that removing Predictive Factor Guidance causes the largest performance drop, including up to +102.98% MSE in running amount, and that disabling the Reliability Boost still degrades performance, including up to +47.02% MSE, especially on perception tasks.

This ablation pattern gives the framework a clear internal hierarchy of importance. Predictive Factor Guidance is identified as the most critical component for enhancing predictive performance, while the reliability mechanism functions as a stabilizer of UrbanInfo. The three-agent design is therefore not merely modular; the components contribute asymmetrically, with factor prioritization acting as the primary driver of gains and conflict-driven re-extraction acting as a secondary but still substantial safeguard.

6. Position within urban multi-agent systems research

The term “Urban-MAS” has a broader meaning in urban analytics than the specific 2025 framework name might suggest. In the wider literature, agent-based modelling is presented as a facet of wider Multi-Agent Systems research that explores the collective behaviour of individual agents and the implications that their behaviour and interactions have for wider systemic behaviour. At the interface of MAS and Urban Analytics, this line of work addresses modelling behaviour, the use of new forms of data, calibration under high uncertainty, real-time modelling, AI techniques, large-scale models, and modelling policy (Malleson et al., 2022).

Against that backdrop, Urban-MAS in the sense of (Lou, 30 Oct 2025) occupies a distinct niche. It is not an urban simulation model in the style of MaaSSim, which reproduces a transport system used by travellers, drivers, and a platform through a single-day, discrete-event simulation built atop SimPy (Kucharski et al., 2020). Nor is it a policy microsimulation like the GAMA-based SARS-CoV-2 MAS for Toronto, which models people, buildings, and roads using rich spatial demographics, hourly mobility data, and daily contagion information (Azeglio et al., 2021). Instead, it is a prediction-oriented LLM-based MAS whose agents orchestrate zero-shot knowledge selection, verification, and inference.

At the same time, Urban-MAS shares an identifiable lineage with more recent language-grounded urban multi-agent systems. MAPUS, for example, models participants as autonomous agents and introduces a coordinator agent for fairness-aware selection and route refinement through language-based negotiation (Guo et al., 25 Mar 2026). This suggests that LLM-based urban MAS research is expanding from simulation and optimization toward architectures in which LLMs act as coordinating, extracting, or negotiating agents over structured urban tasks.

A common misconception is to treat all urban multi-agent systems as synonymous with agent-based simulation. The broader literature shows that urban MAS includes simulation, calibration, decision support, and AI integration (Malleson et al., 2022), while Urban-MAS specifically instantiates a prediction framework for zero-shot human-centered urban tasks (Lou, 30 Oct 2025). The naming overlap is therefore substantive rather than accidental: the specific framework is best understood as one member of a larger family of urban multi-agent approaches, distinguished by its reliance on LLM-mediated decomposition and JSON-structured inference.