New Adapter Embeds Foundation Model Predictions in Choice Models

Foundation models are increasingly accurate in choice prediction tasks, yet their predictions often lack the economic consistency required for real-world applications. Issues include illogical outcomes like increased demand with higher prices, implausible willingness-to-pay estimates, and non-zero probabilities for unavailable options. This research addresses these shortcomings by proposing a novel two-stage adapter. The adapter is designed to integrate the choice probabilities predicted by a foundation model as a precomputed feature within the utility function of a multinomial logit model. In the first stage, the structural coefficients of the multinomial logit are fitted using maximum likelihood, incorporating necessary sign constraints to enforce economic principles. The second stage then freezes these structural coefficients and fits a small neural correction layer that operates directly on the foundation model's predictions. A key theoretical contribution is the proof that this composition precisely preserves the multinomial logit's marginal rate of substitution. This ensures that analytically computable values, such as the value-of-time, become a mathematical guarantee rather than an empirical coincidence. Across three datasets and two foundation models, the adapter demonstrated an average accuracy gain of 6.4 percentage points over the standard multinomial logit, maintained 100% cost monotonicity, and produced value-of-time estimates within established economic ranges for transportation datasets. The system also showed graceful degradation under context restriction, retaining significant accuracy gains even with limited foundation model input.