Causal Reward Alignment for Counterfactual Scenario Generation

Decision-making requires not only predicting outcomes under a given set of conditions, but also reasoning about the scenarios that may arise and how outcomes change under alternative interventions. Tabular data generators provide a natural tool for this task by producing realistic synthetic scenarios and supporting conditional counterfactual queries. However, existing approaches either focus primarily on reproducing observational distributions or impose causal structure through explicit modeling assumptions. We introduce CRAFT, a reward-alignment framework that fine-tunes a queryable autoregressive generator using a dual objective for distributional realism and causal coherence, without encoding a causal graph or structural equations in the generator. Across three semi-synthetic settings and the Twins benchmark, CRAFT substantially improves factual and counterfactual consistency while largely preserving observational realism. Proxy causal rewards estimated from observational data recover much of the benefit of oracle feedback, suggesting a practical route toward causally coherent scenario generation for decision-making.