FairOpt-PFN: Amortized Counterfactual Fairness with Optimal Fair Targets

Ensuring algorithmic fairness is critical for both ethical and legal reasons. Grounded in causal modeling, Counterfactual Fairness provides a framework that aligns well with human intuition. However, traditional approaches to counterfactual fairness require knowledge of the true Structural Causal Model (SCM) at inference time. While much recent work has focused on removing this barrier, Prior-Data Fitted Networks (PFNs) have emerged as a particularly powerful paradigm for amortized bayesian inference. Recent work leverages PFNs to perform counterfactually fair inference using purely observational data, thereby bypassing the need for explicit causal modeling of the fairness task at hand. Yet, the pre-training objective utilized in this approach unnecessarily discards valid, counterfactually fair predictive signal. In this work, we propose a novel pre-training objective derived around utility-optimal decision-making subject to the counterfactual fairness constraint. Evaluating our approach through strictly controlled causal experiments that systematically scale bias, we demonstrate that this pre-training target yields a better fairness-utility trade-off.