Auditing and Fixing Economic Validity in Tabular Foundation Models for Discrete Choice

Researchers propose a two-stage adapter that constrains tabular foundation model predictions within economic theory frameworks, ensuring price-demand relationships remain logically consistent while recovering accuracy gains over standard choice models. The approach achieves up to 13 percentage points of accuracy improvement on transportation datasets while guaranteeing economic validity—a problem raw foundation models fail to solve.

Foundation models have demonstrated impressive predictive accuracy across domains, yet their application to economic choice modeling reveals a critical vulnerability: predictions often contradict basic economic principles. When a price increase correlates with higher predicted demand, or willingness-to-pay estimates turn negative, the models fail at their fundamental task despite statistical performance. This paper addresses a real tension in applying general-purpose AI to specialized domains with embedded logical constraints.

The two-stage adapter framework elegantly solves this by nesting foundation model capabilities within a theoretically grounded utility-maximization model. The first stage establishes economically valid parameters through constrained optimization, while the second stage learns residuals that incorporate foundation model insights without violating constraints. This architectural approach acknowledges that foundation models capture genuine patterns in data but need guardrails to maintain theoretical coherence.

For practitioners deploying AI in economics, transportation, and policy analysis, this work signals that raw foundation model outputs may require domain-specific auditing and correction. The method's ability to maintain perfect economic consistency while improving accuracy over standard logit models represents a meaningful advancement for applied economists, particularly those building pricing engines or demand forecasting systems where economic validity directly impacts business decisions.

The research opens questions about similar constraint-embedding techniques for other specialized domains where foundation models must satisfy domain logic. As foundation models increasingly penetrate professional applications, similar validation frameworks may become standard practice rather than optional refinements.

• →Foundation models achieve high accuracy on choice prediction but frequently violate economic logic like monotonic price-demand relationships.
• →A two-stage adapter embeds foundation model predictions within utility-maximization frameworks to guarantee economic consistency.
• →The method recovers up to 13 percentage points of accuracy improvement over standard logit models while maintaining perfect economic validity.
• →Raw foundation models and conventional distillation fail to achieve both accuracy gains and economic consistency simultaneously.
• →Domain-specific constraint embedding emerges as a necessary practice for deploying general-purpose AI in specialized applications.