CTD decouples convergence then diversity in MOBO, winning 72.9% of 446 comparisons

The arXiv paper 2609.13396 presents converge-then-diversify (CTD) as two sequential phases. Stage one applies standard acquisition functions to drive solutions to a single Pareto point. Stage two switches acquisition functions to expand coverage along the front. Experiments cover DTLZ, WFG and real-world problems under budgets from 50 to 200 evaluations and dimensions up to 20.

Across all 446 comparisons CTD records wins in 325 cases, ties in 94 and losses in 27. The margin widens when evaluation budgets fall below 100 or when objective spaces exceed 8 dimensions. Simple instantiations using expected hypervolume improvement then crowding-distance acquisition already suffice; no new surrogate modelling is required.

Prior MOBO literature treats convergence and diversity as simultaneous objectives inside a single acquisition function. CTD shows this joint optimisation becomes brittle once sample counts are insufficient to populate the entire front. The sequential split removes the need for explicit diversity penalties during early search, reducing hyperparameter sensitivity.

Follow-on work should test CTD on constrained and noisy problems where early convergence may lock into infeasible regions. Integration with batch and parallel acquisition functions remains open.