Embedding Foundation Model Predictions in Discrete-Choice Models with Structural Guarantees

Discrete-choice models are essential for policy decisions, such as calculating the value of time for infrastructure investments or predicting consumer demand. While traditional multinomial logit (MNL) models provide strong structural guarantees—such as cost monotonicity and interpretable willingness-to-pay—they often lack the predictive accuracy of modern tabular foundation models. This paper investigates how to integrate the high predictive power of foundation models into discrete-choice frameworks without sacrificing the economic logic required for reliable policy analysis.

The authors introduce a two-stage adapter that treats foundation model predictions as precomputed features embedded within an MNL utility function.

• Stage 1: The structural coefficients of the MNL are estimated using maximum likelihood with sign constraints, ensuring the model respects economic theory (e.g., higher costs decrease utility).
• Stage 2: These structural coefficients are frozen, and a small neural network (a correction layer) is trained to map the foundation model's probability outputs to a utility adjustment.

This two-stage design is mathematically proven to preserve the marginal rate of substitution (such as the value of time) of the underlying MNL, ensuring that these metrics remain consistent regardless of the neural correction's complexity.

The adapter successfully bridges the gap between predictive accuracy and structural validity. Across three datasets (Swissmetro, LPMC, and IoT-Wearables) and two foundation models (TabPFN and Mitra), the adapter achieved an average accuracy gain of 6.4 percentage points over the baseline MNL. Unlike raw foundation models, which often failed to maintain cost monotonicity or produced implausible willingness-to-pay estimates, the adapter maintained 100% cost monotonicity and produced economically sound trade-off ratios in all tested scenarios. Furthermore, the model demonstrated graceful performance degradation, retaining significant accuracy gains even when the foundation model's context was restricted to 10% of its original size.