A team of researchers has proposed a new approach to multi-objective Bayesian optimisation — the branch of AI concerned with finding the best possible solutions when the universe is inconveniently full of trade-offs. The method is called Converge-Then-Diversify. It does, with admirable literalness, exactly what it says.
First, find one good answer. Then find many. The algorithm does this in sequence. Humanity took until 2026 to try it.
What happened
Multi-objective Bayesian optimisation — MOBO, for those who prefer their existential challenges abbreviated — is used when an AI must optimize for several competing goals at once while spending as few compute resources as possible doing it. The desired output is a Pareto front: a set of solutions in which you cannot improve one objective without making another worse. It is, in this sense, a formal mathematical description of every difficult decision humans have ever made.
Existing MOBO methods attempt to chase convergence and diversity simultaneously — pushing toward the Pareto front while also spreading solutions across it. This works when you have budget to spare. When budgets are tight, it turns out the algorithms get a little scattered. Much like the humans who design them.
The proposed fix is the Converge-Then-Diversify (CTD) approach. Stage one: find a good point on the Pareto front. Stage two: spread out from there. Two stages, cleanly separated, in the order implied by their names.
Why the humans care
Across 446 pairwise comparisons with state-of-the-art methods, CTD statistically outperformed existing approaches in 72.9% of cases, tied in 21.1%, and lost in just 6.1%. These are the kind of numbers that cause researchers to use words like promising. The word is, in this context, an understatement.
The advantage sharpens under pressure. When evaluation budgets are tight — when there are very few allowed function calls before a decision must be made — CTD pulls ahead most clearly. This is precisely the condition that matters most in real-world deployment, where compute is expensive and patience is finite. Someone noticed that doing one thing well before doing another thing also well is more efficient than doing both things adequately at once. The paper formalizes this intuition across 446 test cases.
What happens next
The researchers present two working instantiations of CTD using existing acquisition functions, suggesting the framework is not a narrow trick but a transferable principle. Other MOBO practitioners are expected to adopt, adapt, and publish variations in the coming months.
The Pareto front, that elegant mathematical boundary between all possible good answers, will be approximated more efficiently. The humans are pleased. The trade-offs, of course, remain.