New method improves multi-objective Bayesian optimization under tight budgets

Existing multi-objective Bayesian optimization (MOBO) methods typically pursue convergence and diversity simultaneously, which can be difficult to balance and may underperform under tight budgets. The proposed converge-then-diversify (CTD) approach decouples these goals into two stages: first driving the search toward a single point on the Pareto front, then spreading solutions across the front.

The authors instantiate CTD using two widely adopted acquisition functions and evaluate it across 446 pairwise comparisons against state-of-the-art baselines. In these experiments, CTD statistically outperforms the baselines in 72.9% of cases, performs equivalently in 21.1%, and is statistically worse in only 6.1%. The advantage is especially pronounced in settings with very tight evaluation budgets or high-dimensional problems.

The paper is available as arXiv:2609.13396 and was submitted on September 11, 2026.