CVaR-GPA fine-tunes generative models to capture extreme tail events

A new paper from researchers at UMass Amherst introduces CVaR-GPA (CVaR-penalized Generative Particle Algorithm), a tail-agnostic method for fine-tuning generative models to capture extreme events. Traditional generative models often fail on heavy-tailed distributions because their Lipschitz transport maps preserve the tail behavior of light-tailed source data, causing samples to miss rare but critical events. CVaR-GPA overcomes this by using Wasserstein gradient flows of a Lipschitz-regularized KL divergence penalized by a Conditional Value-at-Risk (CVaR) discrepancy term. The CVaR penalty restores velocity in under-sampled tails, preventing the premature vanishing that plagues standard approaches.

What makes CVaR-GPA particularly practical is its flexibility: it fine-tunes the output samples of any pre-trained generative model without needing access to the model's internal architecture or retraining. It runs on an adaptive time horizon driven by a kinetic-energy stopping criterion, eliminating preset depth tuning. The method derives first-variation subgradients of CVaR using the Rockafellar-Uryasev representation, valid where classical density-based formulas break down. In experiments on synthetic isotropic and anisotropic Student-t distributions, Neal's funnel, and the high-dimensional Fama-French 25 portfolio dataset, CVaR-GPA dramatically improves both global and tail accuracy over baseline pre-trained models, making it a promising tool for risk-sensitive applications in finance, climate science, and reliability engineering.