GDRO framework cuts inventory regret by 60% with Sinkhorn pairing

Generative Distributionally Robust Optimization (GDRO), introduced by Zhang, Li, and Jin, solves a long-standing trade-off in distributionally robust optimization (DRO): existing approaches either accept arbitrary samplers but fail to restrict worst-case distributions, or rely on model-specific access like likelihoods or scores. GDRO's key innovation is a sampler-Sinkhorn pairing: conditional generators are represented exactly via samplers, while Sinkhorn divergence compares their induced distributions using only samples—no likelihoods needed. The population problem admits a differentiable primal-dual formulation at the active decision context, with a finite-sample approximation that works out of the box. For Lipschitz losses, the Sinkhorn radius provides a theoretical guarantee on downstream performance degradation.

In experiments, GDRO reduces rare-context inventory regret by 60% relative to nominal decisions across both explicit and implicit generative models. In SocialGAN navigation, collision rates drop by 50% under distribution shift. The framework is model-agnostic: it accepts any sampleable conditional generator as the nominal, and restricts worst-case laws to a chosen conditional generator family. This opens up DRO to modern generative models (e.g., diffusion models, VAEs, GANs) without requiring internal model access. The authors provide code and a differentiable implementation, making GDRO immediately applicable to supply chain, robotics, and safety-critical ML systems.