Automating MIP Solver Configuration with Large Language Models

Configuring Mixed Integer Programming (MIP) solvers is critical for performance but requires expert knowledge and extensive tuning. We show that large language models (LLMs) discover strong configurations by iteratively proposing settings and refining them from solver feedback, matching or exceeding Bayesian optimization and Gurobi's autotuner per instance. Aggregating these configurations yields a single per-class configuration that reduces the primal-dual integral by 32.0\%-86.0\% over Gurobi defaults and 18.8\%-52.1\% over SCIP defaults on five classical problem classes, deployable as a drop-in replacement for solver defaults.