Bridging Spherical Black-Box Optimizers

A team of researchers has unified several major black-box optimization methods under a single theoretical framework, identifying two core design choices that distinguish them and using that insight to build hybrid optimizers that outperform their predecessors on standard benchmarks and higher-dimensional locomotion tasks. The work, submitted in June 2026 by Johannes Ackermann and Stefano Peluchetti, addresses a fundamental challenge in machine learning: optimizing systems when gradient information is unavailable [1][3]. In such settings, practitioners turn to black-box optimization (BBO) methods. Three prominent families—Evolution Strategies (ES), Consensus-Based Optimization (CBO), and Optimization via Integration (OVI)—have historically been studied in isolation [1][2]. The new paper reveals these approaches differ primarily along two axes: fitness aggregation, which controls a preference for sharpness, and consensus scope, which controls modality [1][3]. By making these design choices explicit, the authors introduce hybrid optimizers that interpolate between existing methods. An ES-OVI hybrid allows explicit control over the preference for flat minima, enabling a trade-off between performance and robustness in continuous control tasks [1][2]. A CBO-OVI hybrid combines the higher-dimensional efficiency of parametric methods with the multimodal capabilities of particle-based approaches, achieving competitive results on language model merging under limited evaluation budgets [1][3]. The hybrid methods were validated on standard BBO benchmarks and higher-dimensional locomotion tasks, where they outperformed their constituent algorithms [1][2]. The broader BBO landscape has seen parallel efforts to improve solver selection. A separate 2025 study introduced a geometric probing framework that represents problem instances by randomly sampled multi-scale two-dimensional slices of the objective landscape, encoding them with validity-mask-aware visual pooling to retain local patterns such as basin shape and anisotropy [4]. Another line of work from 2025 proposed a greedy restart scheduling approach that creates a static solver schedule from a database of previously evaluated optimizers, closing the gap from the single best solver to the virtual best solver by more than 95% regarding relative ERT on the BBOB testbed [5].