💻 computer science

CauCity-WM: An Intervention-Consistent Generative World Model for Counterfactual Urban Dynamics and Risk-Aware Decision Making

The paper introduces CauCity-WM, a causal generative world model that unifies urban forecasting, counterfactual estimation, and risk-aware decision making by employing a twin-world structural diffusion process to generate consistent factual and counterfactual futures through abducted exogenous noise.

Original authors: Kai Ma, Miao Yu

Published 2026-09-10

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Kai Ma, Miao Yu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Cities are living, breathing systems of immense complexity. Every day, millions of vehicles move along a web of roads, reacting to traffic lights, weather, accidents, and the decisions of drivers. For decades, scientists and planners have built computer models to predict how this traffic will flow. These models are excellent at answering the question, "What will happen next?" based on what is happening right now. They can forecast a traffic jam ten minutes from now with impressive accuracy. However, city planning often requires a different kind of question, one that is much harder to answer: "What would have happened if we had made a different choice?" If a city planner had closed a specific lane yesterday, or changed the timing of a traffic signal, how would the traffic pattern have shifted? This is the realm of counterfactual reasoning—imagining a different reality for the same specific moment in time. The challenge is that in the real world, we cannot rewind the clock and run the same day twice with different rules. We only have one version of history, and the random events that occurred, like a sudden rainstorm or a minor fender bender, are unique to that single timeline.

A team of researchers has introduced a new approach to this problem called CauCity-WM. It is a type of computer model designed to simulate urban dynamics in a way that respects the unique randomness of each specific day. Unlike previous models that might generate a new, random future every time they are asked to imagine a different scenario, this system is built to keep the underlying "noise" of the day constant. Imagine a city as a complex machine where the weather, the time of day, and the general mood of the drivers create a specific background hum. When the researchers ask, "What if we changed the traffic lights?", their model does not change that background hum. It keeps the exact same weather, the exact same minor incidents, and the exact same random fluctuations in driver behavior, and only swaps out the traffic light setting. By holding everything else steady, the model can isolate the true effect of the change, showing exactly how the city would have responded to that specific intervention without the confusion of new random events clouding the result.

The researchers tested this idea using real-world data from seven different urban forecasting benchmarks, which included traffic speed and flow data from thousands of sensors across various cities. They compared their new model against the strongest existing forecasting tools and specialized causal models designed to handle "what if" questions. The results showed that while the new model was just as good at predicting normal traffic as the best existing tools, it was significantly better at answering counterfactual questions. When asked to estimate the effect of a hypothetical action, the new model reduced the average error in its predictions by nearly 19 percent compared to the next best method. It also achieved a higher accuracy in determining whether an action would increase or decrease traffic congestion, getting the direction of the effect right more than 84 percent of the time. Furthermore, the model was able to generate realistic confidence intervals, meaning it could tell planners not just what would happen, but how sure it was about that prediction, covering the true outcome 89.6 percent of the time in its tests.

To ensure these findings were not just mathematical tricks, the researchers also tested the model in a highly detailed microscopic traffic simulator called SUMO. In these simulations, they created controlled scenarios where they could compare the model's advice against a known reality. They asked the model to help manage traffic by adjusting signal timings, ramp meters, and lane capacities. When the model's risk-aware planning was used, the average travel time for vehicles dropped to 27.1 minutes, a significant improvement over other methods. More importantly, the model drastically reduced the number of times traffic rules were violated or queues grew dangerously long, cutting constraint violations down to just 1.9 percent. This suggests that the model does not just predict the average outcome well; it also understands the rare, extreme events that cause gridlock, allowing planners to avoid them.

The core innovation behind this success is the way the model handles the concept of "twin worlds." In many computer simulations, when you ask for a different outcome, the computer generates a completely new set of random numbers to create that new world. This makes it impossible to tell if the difference in the result is due to the change in action or just the new randomness. The CauCity-WM model avoids this by using a "noise-preserving" technique. It first looks at the real, observed day and figures out the specific random disturbances that happened. Then, when it simulates a counterfactual day, it uses those exact same disturbances. It is like watching a movie and then asking, "What if the hero had taken a different path?" The new model re-runs the scene with the same weather, the same background noise, and the same audience reactions, changing only the hero's path. This allows the researchers to see the pure effect of the decision, stripped of the confusion caused by changing the environment.

The researchers also found that this approach helps the model work across different cities and situations. By aligning the way the model learns from different environments, it can transfer its knowledge to cities it has never seen before. When tested on a massive dataset covering thousands of sensors across California, the model maintained its accuracy even when asked to predict traffic in regions it had not been explicitly trained on. This is crucial for practical application, as city planners often need to make decisions for areas where historical data is scarce. The model also demonstrated robustness when faced with unseen types of interventions, such as new combinations of traffic management strategies, proving that it learned the underlying rules of traffic flow rather than just memorizing specific examples.

Despite these successes, the researchers are careful to note the limits of their work. The model relies on the assumption that it can see enough of the past to understand the hidden factors driving the traffic. If there are critical pieces of information missing, such as an unrecorded police operation or a sudden shift in public sentiment, the model's counterfactual predictions could be biased. The researchers stress that their model is a tool for exploring possibilities, not a crystal ball that can predict the future with absolute certainty. It is most effective when used to evaluate actions that are within the range of what has been observed before, rather than extreme, unprecedented scenarios like a city-wide evacuation. The study concludes that while the model represents a significant step forward in understanding urban dynamics, it is part of a larger toolkit that should include human judgment and other data sources.

The implications of this work extend beyond just traffic. By creating a system that can reliably simulate the consequences of different actions in a complex, noisy environment, the researchers have provided a new way to think about decision-making in cities. Whether it is planning a new bus route, managing energy consumption, or responding to a public health crisis, the ability to ask "what if" with confidence is invaluable. The CauCity-WM model shows that it is possible to build computer systems that respect the complexity of the real world while still offering clear, actionable insights. It moves the field of urban planning from simply predicting the future to actively exploring and shaping it, offering a practical foundation for making decisions that are not only smart but also safe and risk-aware.

Technical Summary: CauCity-WM

Problem Statement

Urban decision-making requires more than accurate observational forecasting; it necessitates the ability to estimate how a realized episode would evolve under an action that was not taken (counterfactual reasoning). Existing urban forecasting models, including graph recurrent networks, Transformers, and urban foundation models, are optimized primarily for observational prediction. While they capture evolving correlations and exogenous contexts effectively, their learned transition dynamics are generally not defined as structural intervention operators. Consequently, high predictive fidelity does not guarantee that the model can determine what a specific episode would have become under a do(A) intervention.

The core challenges identified are:

Stochastic Intervention Consistency: At the episode level, factual and counterfactual worlds must share the latent disturbances that generated the realized episode. Independent sampling entangles treatment effects with random realizations (e.g., weather, incidents, sampler noise).
Partial Observability: The relevant structural disturbance is latent. Valid counterfactuals depend on the ability to abduct this disturbance from sufficient history and ensure the intervention remains within a supported action region.
Spatial Locality and Cross-Regime Stability: Models must allow intervention effects to propagate through plausible graph paths while avoiding unrestricted global responses, and must preserve transferable mechanisms without erasing city-specific heterogeneity.
Risk-Aware Planning: Decision-making requires calibrated outcome distributions, as mean-optimal actions may lead to unacceptable tail congestion or constraint violations.

Methodology: CauCity-WM

CauCity-WM is a Causal Generative Urban World Model built around a noise-preserving intervention operator for graph-conditioned diffusion dynamics. Unlike approaches that treat causality as an auxiliary prediction head, CauCity-WM embeds intervention semantics directly into the stochastic transition law.

Core Architecture

The model operates on a latent structural state $Z_t$ and an episode-level exogenous disturbance $U_t$ . The transition is a reparameterized stochastic map:
$Z_{t+1} = D_\theta(Z_t, A_t, E_t, G_t, U_t; \epsilon_{t,1:K})$
where $D_\theta$ is a graph-conditioned diffusion law, $A_t$ is the action, $E_t$ is exogenous context, $G_t$ is the urban graph, and $\epsilon$ is base diffusion noise.

Key Mechanisms

Noise-Preserving Twin-World Diffusion:
- Abduction: For a realized factual trajectory, the model abducts an episode-specific structural disturbance $\hat{U}$ and matches the diffusion base-noise path $\hat{\epsilon}$ .
- Twin-World Rollout: To generate a counterfactual, the model reuses the same $\hat{U}$ and $\hat{\epsilon}$ but substitutes the action $A_t$ with the intervention $\tilde{A}_t$ .
- Semantics: This creates paired factual and counterfactual worlds at the level of the learned transition, ensuring that the only difference between the two worlds is the intervention. This reduces Monte Carlo variance in effect estimation (Proposition 2) and provides unit-level counterfactual semantics.
Structured Latent State & Graph Dynamics:
- Historical observations are encoded via a causal temporal encoder and topology-aware aggregation using graph attention.
- The latent state $Z_t$ is optimized jointly with structural transition and counterfactual objectives to support stable action-response mechanisms.
Mechanism Invariance and Transfer:
- Gradient Alignment: A mechanism-gradient alignment loss ( $L_{grad}$ ) and residual distribution matching ( $L_{res}$ ) constrain the transition law to be stable across different cities and regimes (environments), separating shared mechanisms from environment-specific contexts.
- Sparse Intervention Jacobians: A regularizer ( $L_{act}$ ) enforces that actions propagate only through supported graph paths, preventing unrestricted global responses and encoding action-locality.
Unified Inference Modes:
- Factual Forecasting: Standard prediction using the transition law.
- Retrospective Counterfactuals: Uses the observed future to abduct $U$ and evaluate "what if" scenarios.
- Prospective Planning: Uses a history-conditioned prior for $U$ to optimize action sequences.
- Risk-Aware Decision Making: Optimizes a CVaR-regularized objective over the sampled transition distribution to balance utility (e.g., travel time) against tail risks (e.g., congestion, constraint violations).

Key Contributions

Noise-Preserving Intervention Operator: Defines a stochastic intervention operator for graph-conditioned diffusion where episode-level structural disturbances and base noise are held fixed across twin worlds. This links the construction to Structural Causal Model (SCM) abduction–action–prediction under explicit assumptions.
Action-Local and Transferable Structural Dynamics: Constrains the transition law with mechanism-gradient alignment and sparse intervention Jacobians, characterizing local propagation through supported graph paths.
Unified Framework: Integrates factual forecasting, retrospective counterfactual analysis, prospective intervention evaluation, and CVaR-regularized planning into a single transition distribution, avoiding separate training of treatment-effect or policy networks.
Expanded Validation: Provides extensive evaluation across seven real-world forecasting benchmarks, semi-synthetic counterfactuals, adapted causal baselines, and calibrated SUMO simulation scenarios.

Experimental Results

The model was evaluated on seven real-world datasets (METR-LA, PEMS-BAY, PEMS04, PEMS08, LargeST, NYC TLC, Citi Bike) and a family of calibrated SUMO scenarios.

Factual Forecasting: CauCity-WM achieved the lowest Mean Absolute Error (MAE) across all seven datasets compared to strong baselines (e.g., OpenCity, UniST, STAEformer), demonstrating that adding intervention semantics does not sacrifice predictive accuracy.
Counterfactual Fidelity: In counterfactual evaluation, CauCity-WM (twin-noise) achieved an Effect-MAE (E-MAE) of 2.61 and PEHE of 3.29, outperforming the strongest adapted causal baseline (Diﬀ-SCM, E-MAE 3.22, PEHE 4.07). It also achieved 84.6% effect-sign accuracy and 89.6% coverage of nominal 90% effect intervals.
Transferability: In zero-shot transfer tests on held-out regions (LargeST), CauCity-WM showed the smallest relative degradation (+22.1%) compared to other models, attributed to mechanism invariance.
Decision Quality: In calibrated SUMO scenarios, CauCity-WM reduced mean travel time to 27.1 minutes (vs. 29.8 for DreamerV3) and constraint violations to 1.9% (vs. 4.1%), while significantly lowering congestion CVaR.
Ablation Studies: Removing twin-noise coupling significantly degraded counterfactual fidelity (PEHE increased from 3.29 to 4.31) with minimal impact on factual MAE, confirming the specific value of the noise-preserving mechanism. Removing mechanism invariance significantly hurt zero-shot transfer performance.

Significance and Claims

The paper claims that intervention-consistent urban world modeling provides a practical basis for counterfactual analysis and risk-aware decision making. The primary significance lies in shifting the paradigm from treating causality as an auxiliary task to embedding stochastic intervention consistency directly into the generative transition.

The authors emphasize that:

The "noise-preserving" mechanism isolates action-induced changes from episode-irrelevant stochastic variation, directly addressing the variance problem in counterfactual estimation.
The model does not claim non-parametric identification from observational data alone; rather, it makes explicit the assumptions (state sufficiency, posterior recoverability, positivity, mechanism stability) required for valid counterfactuals.
The approach enables a single structural generative model to support forecasting, counterfactual reasoning, and constrained planning, offering a unified computational laboratory for urban planners to ask "what would happen if..." questions with calibrated uncertainty.

The paper concludes that while partial observability and unsupported interventions remain limitations, the proposed design principle of noise-preserving intervention consistency offers a distinct and effective path forward for urban intelligence.