New arXiv paper enables risk-aware AI without extra cost

In high-stakes domains like healthcare and finance, AI systems must often learn from logged data (offline) and carefully control adverse outcomes. Standard contextual bandit algorithms typically focus on maximizing expected reward, ignoring risk. This new work, submitted to arXiv on May 15, 2026, introduces a pessimistic risk-aware policy learning framework that optimizes a broad class of risk measures—including conditional value-at-risk (CVaR), mean-variance, and entropic risk—without requiring online interaction.

The key theoretical contribution is a set of novel empirical concentration inequalities for importance sampling-based distributional estimators. These allow the authors to derive data-dependent suboptimality bounds at a rate of $\tilde{\mathcal{O}}(1/\sqrt{n})$, which is minimax optimal and matches the rate of risk-neutral offline policy optimization. Importantly, this rate holds without the restrictive uniform overlap assumption commonly needed in causal inference. The result implies that optimizing general Lipschitz risk criteria incurs no additional statistical cost relative to expected-reward optimization, opening the door to safer AI deployment in high-stakes offline settings.