SPARROW Optimizes Black-Box Functions with Low Budgets

Black-box optimization is a fundamental technique for optimizing objectives without gradient information, but it often requires numerous costly function evaluations, especially when dealing with noisy or failure-prone evaluations and complex search spaces. Existing generative model-based methods for navigating these spaces typically need a large number of evaluations to align their samplers effectively, making them impractical in low-budget scenarios. To overcome this, researchers propose SPARROW, an algorithm that completely separates the generative prior from the reward signal. SPARROW can utilize any sampler with a known corruption process, trained on unevaluated data, as a fixed, structured proposal operator. Optimization then proceeds through rank-based guidance over an archive of evaluated candidates. This approach allows SPARROW to navigate intricate geometries, manage unreliable reward signals, and perform effective optimization even with very limited evaluation budgets, offering asymptotic convergence guarantees over the sampler support.