Conformal Policy Learning with Distribution-Free Safety Guarantees
The paper proposes a treatment-assignment method that limits the chance of giving treatment to someone it would harm.
CPL frames each decision as a test for counterfactual harm, using conformal p-values and selective calibration when both potential outcomes cannot be observed. For randomized experiments, the authors claim finite-sample safety at a user-chosen level without outcome-modeling assumptions. They also report asymptotically optimal welfare under a consistently estimated outcome model, and doubly robust safety guarantees for observational studies using learn-then-balance weights. The method is tested in simulations and applied to AI-powered interventions aimed at reducing conspiracy beliefs. ArXiv · AI/CL/LG's note
CPL frames each decision as a test for counterfactual harm, using conformal p-values and selective calibration when both potential outcomes cannot be observed. For randomized experiments, the authors claim finite-sample safety at a user-chosen level without outcome-modeling assumptions. They also report asymptotically optimal welfare under a consistently estimated outcome model, and doubly robust safety guarantees for observational studies using learn-then-balance weights. The method is tested in simulations and applied to AI-powered interventions aimed at reducing conspiracy beliefs. ArXiv · AI/CL/LG's note
score 5