Log-Driven Semantic Tuning for Mixed Integer Programming Solvers

Configuring mixed-integer linear programming (MILP) solvers is costly when each candidate requires a solver run. Scalar outcomes rank configurations, but execution logs also reveal where the solver spends effort. We introduce Log-Driven Semantic Tuning (LDST), in which a large language model interprets these diagnostics using parameter meanings and a bounded history of trials. The model proposes changes to the best configuration found so far, automating the interpretation and editing performed between runs. Across 215 MILP instances evaluated with SCIP, LDST reduces geometric mean best-so-far loss by 43.9% and 16.7% relative to the best black-box baseline in the easy and intermediate regimes after 50 proposals. The benchmark protocol and tuning trajectories provide a common basis for studying how semantic knowledge and execution feedback guide configuration.