Bridging Spherical Black-Box Optimizers

Abstract: When gradient information is unavailable, black-box optimization (BBO) methods provide a practical alternative. While Evolution Strategies (ES), Consensus-Based Optimization (CBO), Optimization via Integration (OVI), and related methods have each been studied independently, their connections remain underexplored. We unify these approaches within a common theoretical framework, revealing that they differ primarily in two design choices: fitness aggregation (controlling sharpness preference) and consensus scope (controlling modality). Leveraging these insights, we introduce hybrid optimizers that interpolate between existing methods. Our ES-OVI hybrid allows explicit control over the preference for flat minima, enabling a trade-off between performance and robustness in continuous control tasks. Our CBO-OVI hybrids combine the higher-dimensional efficiency of parametric methods with the multimodal capabilities of particle-based approaches, achieving competitive results on LLM merging under limited evaluation budgets. We validate our methods on standard BBO benchmarks and higher-dimensional locomotion tasks, demonstrating that the hybrid methods can outperform their constituent algorithms.

• The paper unifies spherical black-box optimizers by bridging ES, OVI, and CBO through a master update equation.
• It introduces hybrid methods like ES-OVI and cCBO-OVI that balance sharpness and robustness without extra computational cost.
• Experimental results on BBOB and Brax benchmarks demonstrate superior performance for hybrid methods in multi-modal, high-dimensional tasks.

Bridging Spherical Black-Box Optimizers: A Unified Framework

Introduction and Motivation

"Bridging Spherical Black-Box Optimizers" (2606.25761) offers a rigorous unification of core families of spherical black-box optimization (BBO) methods—Evolution Strategies (ES), Optimization via Integration (OVI), and Consensus-Based Optimization (CBO), including higher-order variants (pCBO, cCBO, DE). The motivation is the fragmented progression of these techniques: despite structural similarities, their connections and possible hybridizations remain under-explored. Unifying them serves to demystify their foundational algorithmic tradeoffs and yields new classes of optimizers, capable of exploiting the synergies of their parent methodologies. The resulting framework provides explicit control over optimization characteristics such as sharpness preference and modality—essential in settings from continuous control to non-convex, multi-modal model merging.

Unified Master Update Formalism

Central to the contribution is a master update equation parameterized in terms of two orthogonal axes:

• Fitness Aggregation: Governs sharpness preference, ranging from linear averaging (ES) to log-sum-exp (OVI).
• Consensus Scope: Regulates interaction, covering global consensus (ES, OVI, CBO) to local (pCBO) or clustered consensus (cCBO).

This formulation asserts that the vast majority of popular spherical BBO methods are differentiated only by choices for these axes—specifically, the fitness transformation $\Psi(F)$ and the interaction kernel $K_t^{i,j}$.

(Figure 1)

Figure 1: ES objective favors broad, flat optima (robustness), while OVI maintains sharp minima (performance); aggregation and scaling pivot the optimizer's convergence preferences.

Analysis and Synthesis of Algorithms

ES, OVI, CBO: Comparative Analysis

ES performs mean-shifting on a Gaussian-perturbed search distribution, minimizing the expected loss. In contrast, OVI optimizes a soft-min (log-sum-exp) surrogate, yielding higher selectivity for sharp minima. CBO and its nonparametric variants propagate populations toward a weighted consensus determined by fitness, either globally or through local kernels.

(Figure 2)

Figure 2: The ES-OVI hybrid introduces a tunable $\alpha$ allowing for smooth interpolation between flat (robust, ES-like) and sharp (high-performing, OVI-like) optima on the Rosenbrock function.

Notably, OVI formalizes a distributional update that is equivalent to Consensus Hopping (CH), a type of CBO with global consensus and constant noise. This equivalence establishes a key theoretical bridge: parametric and particle-based approaches can be aligned under shared update dynamics.

Hybridization: Interpolation and Adaptation

ES-OVI: Gradient Interpolation

Formally, the ES-OVI hybrid interpolates between ES and OVI gradients:

$$\nabla J^\alpha(\theta) = \alpha \nabla J^\mathrm{OVI}(\theta) + (1-\alpha) \nabla J^\mathrm{ES}(\theta), \qquad \alpha \in [0,1].$$

This does not increase per-iteration objective evaluations, enabling per-task optimizer tuning—balance between robustness and sharpness—without increased computational expense.

Multi-Modal Adaptivity: cCBO-OVI Hybrids

Particle-based approaches such as cCBO and DE are required for multi-modal optimization in high dimensions (e.g., evolved model merging). OVI is efficient in high-$d$ due to its parametric nature but is unimodal by design. By devising AdaPol (adaptive interpolation) and SchedPol (scheduled consensus transition), the authors fuse cCBO's multi-modality with OVI's scalability. The adaptive method allocates the particle population dynamically between OVI-like and cluster-based updates based on recent optimization success.

(Figure 3)

Figure 3: Model merging under evaluation constraints is explicitly multi-modal—multimodal optimizers (cCBO, SchedPol) can discover diverse optima, outperforming unimodal methods in regimes with good initialization.

Experimental Results

BBOB and Brax Benchmarks

Experiments span:

• BBOB: Low-dimensional, multi-modal synthetic problems, demonstrating previously unreported results for CBO and OVI on this established suite.
• Brax: High-dimensional RL tasks (MLP with $d \approx 1000$) evaluative of scalability and optimizer robustness.

(Figure 4)

Figure 4: On BBOB 2D tasks, ES-OVI hybrids can outperform either parent method, particularly on functions sensitive to the sharpness/robustness tradeoff.

(Figures 10, 11, 12, 13, 14)

Figures 10–14: Higher-dimensional BBOB tasks (up to 20D) confirm persistence of optimizer trends, but also highlight the dimension-dependent breakdown of purely particle-based methods.

The key empirical findings are:

• ES-OVI advantages exceed both constituent methods in cases where the optimal trade-off between solution sharpness and robustness is problem-dependent.
• AdaPol and SchedPol surpass pure cCBO or OVI in high-dim model merging tasks under tight evaluation budgets, uniquely combining ability to escape local basins (OVI stage) and explore multiple optima (cCBO phase).
• Strong numerical results: For several tasks, the ES-OVI hybrids yield lower final fitness than ES or OVI alone, demonstrating non-trivial synergies not attainable by simple ensembling.

Theoretical Implications and Extensions

The unifying framework presented positions spherical BBO methods within a concise algebraic space of update rules, clarifying the effect of consensus and aggregation choices.

• Parameter Sensitivity Control: Tunable interpolation in ES-OVI provides a direct lever for obtaining robust (flat) versus high-performance (sharp) solutions, a crucial tradeoff in reinforcement learning and generalization-sensitive tasks.
• High-$d$ Multi-Modality: Adaptive and scheduled consensus transitions extend the multi-modal capacity of particle methods to practically relevant high-dimensional spaces.

Limitations arise from the focus on spherical (fixed-covariance) sampling; full covariance adaptation (as in CMA-ES) is orthogonal and cannot be captured by this framework. Additionally, the performance of AdaPol and SchedPol depends on problem-specific parameter schedules or meta-heuristics (e.g., $N_\mathrm{G}$, $\alpha$), whose optimal settings remain an open subject for meta-optimization or learning.

Practical Implications and Prospects

This unifying BBO lens has direct implications for areas where gradient access is infeasible, and the optimization landscape is rugged or multi-modal, such as:

• Reinforcement learning with stochastic/delayed rewards
• Neural policy search in simulators/APIs
• Large model parameter merging with minimal supervision

Future work may extend this analysis to non-spherical or adaptive-covariance regimes, invoke meta-learned update rules to optimize the consensus/aggregation interpolation over task classes, and explore hybridization with advanced population heuristics (e.g., SVGD, PSO) with theoretical guarantees.

Conclusion

The paper establishes that parametric and non-parametric black-box optimizers for spherical problems can be systematically unified and hybridized by dissecting their aggregation and consensus construction mechanisms. The resulting spectrum of optimizers exposes non-trivial tradeoffs (sharpness, multimodality, sample efficiency) that prior art could not simultaneously address. These insights are directly actionable for both practitioners selecting optimizers and for the development of theoretically sound, problem-adaptive BBO algorithms.