New DACMO framework uses LLMs to auto-design optimization algorithms

A team led by Zhiyuan Wang et al. has published a paper proposing DACMO (Domain-Agnostic Co-evolution of Parameterized Search for Multi-Objective Binary Optimization), the first general-purpose framework for constructing parallel algorithm portfolios (PAPs) for multi-objective binary optimization problems (MOBOPs). The key challenge is that current PAP construction methods require problem-specific instance generators or manual tuning, limiting their applicability. DACMO solves this with two technical innovations.

First, the authors introduce a neural instance representation architecture that separates domain-invariant features from instance-specific ones. This allows the framework to generate class-consistent problem instances across varying dimensions without needing manually designed generators for each problem domain. Second, DACMO incorporates LLM-based automatic search operator generation into the PAP construction pipeline, expanding the search space from simple parameter tuning of predefined templates to full operator-level algorithm design. This means the system can create novel search strategies rather than just adjusting existing ones.

Evaluated on four diverse MOBOP classes—multi-objective max match, knapsack, contamination control, and complementary influence maximization—DACMO was applied directly without modification. It outperformed PAPs built from classic multi-objective evolutionary algorithm (MOEA) templates across all four classes. Furthermore, it achieved performance comparable to a state-of-the-art baseline that relies on manually designed problem-specific instance generators, and actually surpassed that baseline on two of the four problem classes. The work represents a significant step toward fully automated algorithm design for complex real-world optimization tasks.