Pareto-optimal estimation and policy learning for balancing short-term and long-term outcomes

In high-stakes domains such as precision medicine and personalized recommendation, optimizing treatments or policies demands a careful balance between immediate rewards and long-term outcomes. However, these objectives often exhibit inherent tradeoffs - for instance, aggressive medication dosages may accelerate short-term recovery yet induce severe long-term side effects. Existing methods struggle to address such dilemmas due to two primary challenges: the lack of explicit mechanisms to reconcile conflicting outcomes and the gradient interference inherent in joint optimization. To address these issues, we propose a Pareto-efficient framework comprising Pareto-Optimal Estimation (POE) and Pareto-Optimal Policy Learning (POPL). Specifically, POE employs a continuous Pareto optimization module to explicitly resolve task-level conflicts between representation learning and multi-outcome prediction. Building upon these counterfactual estimates, POPL explores the treatment-response surfaces over continuous dosage spaces and identifies the Pareto frontier, thus enabling proactive and balanced decision-making. Extensive experiments on both synthetic benchmarks and real-world datasets demonstrate that our approach consistently achieves a superior performance in terms of counterfactual prediction and policy learning.