Three-in-One World Model: Energy-Based Consistency, Prediction, and Counterfactual Inference for Marketing Intervention

Technical Summary: Three-in-One World Model for Marketing Intervention

1. Problem Statement

Marketing decisions involve complex interactions between latent consumer heterogeneity, time-varying internal states, and explicit interventions. Current prediction-oriented models (e.g., standard MLPs) and language-oriented models (e.g., LLMs) fail to capture this structure in a unified manner. Specifically:

• LLMs rely on next-token prediction, lacking explicit mechanisms for causal reasoning or internal world representation.
• Existing World Models in marketing (e.g., DBM-based approaches) have been limited to static co-occurrence, unable to model time-varying dynamics or evaluate the consistency of temporal trajectories.
• Causal Inference Methods (e.g., S-, T-, X-, DR-learners, Causal Forests) often struggle to recover heterogeneous treatment effects (CATE) when treatment assignment is confounded with latent consumer traits, as they typically estimate outcomes rather than isolating treatment effects through counterfactual contrasts.

The paper addresses the need for a unified framework that supports consistency evaluation, outcome prediction, and counterfactual inference within a single architecture capable of disentangling latent traits.

2. Methodology

The authors propose a Three-in-One World Model architecture based on a Deep Boltzmann Machine (DBM) combined with lightweight, task-specific adapters.

Architecture Components:

• World Model (Frozen DBM): A Bernoulli–Bernoulli DBM with stacked hidden layers (64–32–16) serves as the core.
  • Visible Layer: Contains static demographics, calendar features, and lagged time-varying states (recent visit/purchase recency, cumulative counts) and lagged actions/outcomes (past 4 periods). Crucially, the current-period action vector is excluded from the visible layer.
  • Training: The DBM is trained in two stages: greedy layer-wise pretraining of RBMs followed by joint fine-tuning using persistent contrastive divergence.
  • Belief Representation: After training, the DBM is frozen. The "belief" vector is formed by concatenating the mean-field activations of all hidden layers. This representation captures the consumer's state independent of the current intervention.
• Task-Specific Adapters: Lightweight MLPs are trained on top of the frozen belief.
  • Input: The concatenation of the frozen belief vector and the current-period action vector.
  • Function: These adapters map the state-action pair to downstream outcomes (e.g., visit or purchase probability).

Three Integrated Tasks:

1. Energy-Based Consistency Evaluation: The DBM's free energy ($F(v)$) scores the plausibility of an input trajectory. Low energy indicates a plausible configuration of demographics, history, and behavior; high energy flags implausible scenarios.
2. Outcome Prediction: Adapters predict outcomes (visit/purchase) based on the belief state and current action.
3. Counterfactual Inference: To estimate the Conditional Average Treatment Effect (CATE), the belief vector is held fixed (representing the consumer's state), and the current action input to the adapter is toggled (e.g., from 0 to 1). The difference in predicted outcomes yields an individual-level treatment effect estimate.

3. Experimental Setup

The model was evaluated using a controlled simulation where ground-truth latent traits were known:

• Data: 1,024 synthetic consumers over 365 days.
• Latent Parameters: Each consumer had known price sensitivity ($\alpha$), promotion responsiveness ($\gamma$), and base preference ($\beta$).
• Confounding: $\alpha$ and $\beta$ were negatively confounded via income/age; $\gamma$ and $\beta$ were positively confounded via loyalty.
• Baselines: The proposed "Belief Adapter" was compared against:
  • A strong MLP baseline (trained end-to-end on raw features).
  • Meta-learners: S-, T-, X-, and DR-learners.
  • Causal Forest.

4. Key Results

A. Energy-Based Consistency

• The DBM successfully penalized counterfactual trajectories that were implausible given the consumer's history. Specifically, clamping a trajectory to show "purchases without recent promotional exposure" resulted in a significantly higher free energy (lower plausibility) compared to observed trajectories.
• The penalty ($\Delta F$) was heterogeneous: consumers with high latent base preference ($\beta$) received a smaller penalty than those with low $\beta$, demonstrating that the energy landscape reflects latent traits learned from demographics.

B. Prediction Performance

• The "Belief Adapter" matched the performance of the end-to-end MLP baseline on visit and purchase prediction (AUC differences were within $\pm 0.001$). This confirms that the DBM's dimensionality reduction did not sacrifice predictive power.

C. CATE Recovery (Counterfactual Inference)

• Unconfounded Intervention (Push $\to$ Visit): The Belief Adapter achieved the highest Spearman rank correlation ($\rho = 0.625$) between estimated CATE and the true promotion responsiveness ($\gamma$), outperforming the best meta-learner baseline ($\rho = 0.579$).
• Confounded Intervention (Sale $\to$ Purchase): The advantage was most pronounced here. The Belief Adapter recovered the true price sensitivity ($\alpha$) with $\rho = 0.548$. In contrast, standard baselines struggled significantly:
  • T-learner: $\rho = 0.161$
  • Causal Forest: $\rho = 0.040$
  • X-learner: $\rho = 0.123$
  • DR-learner: $\rho = -0.016$
• The adapter also exhibited the expected negative correlation with the confounding parameter ($\beta$), suggesting it relied on the correct latent signal rather than shortcuts.

5. Significance and Claims

The paper claims that the proposed architecture provides a unified substrate for marketing intervention that overcomes specific limitations of current approaches:

• Disentanglement: The DBM belief representation successfully disentangles latent traits (price sensitivity, promotion responsiveness, base preference) in a form that survives counterfactual queries.
• Modularity: By freezing the world model and training only lightweight adapters, the framework avoids catastrophic forgetting when adding new tasks.
• Causal Robustness: Unlike standard meta-learners that operate on raw features, the "frozen belief + varying action" mechanism allows for more faithful recovery of heterogeneous causal effects, particularly in scenarios with confounded treatment assignment.
• Temporal Consistency: Unlike previous static DBM applications, this model evaluates the consistency of time-extended trajectories, not just static co-occurrences.

Limitations Acknowledged by the Authors:

• Results are based on a single simulation seed and data-generating process; replication is needed to characterize variance.
• The energy-based analysis examined a single counterfactual clamp configuration.
• The study is restricted to a single product category and binary outcomes; extension to continuous outcomes and real A/B-test data is future work.
• The paper does not claim to solve optimal policy learning (closing the loop with a controller) but identifies this as a necessary next step.