MILP-Evo Automates Solver Design with LLM-Guided Program Evolution.

Researchers have introduced MILP-Evo, a novel framework for the automatic design of Mixed-Integer Linear Programming (MILP) solver logic. Unlike many machine learning approaches that produce opaque policies, MILP-Evo focuses on generating explicit, white-box solver components that are easy to understand, modify, and integrate into standard solver workflows. The system operates in a closed-loop manner, leveraging large language models (LLMs) to guide the search for optimal solver logic. It iteratively proposes candidate programs, loads them into a solver like SCIP, and evaluates their performance on MILP instances. Feedback from these evaluations then informs performance-based selection, targeted repairs, and diversity maintenance within the population of candidate solutions. Experiments across various benchmark families demonstrate that LLM-guided program evolution can discover competitive, domain-specialized policies. This approach offers a significant step towards automating the design of complex optimization solver components, potentially accelerating the development of more efficient and tailored MILP solutions.