Large Language Models for Solver Selection: A Preliminary Study

This thesis investigates the feasibility of employing Large Language Models (LLMs) as decision components for solver selection in Constraint Programming (CP). Solver selection is a well-known instance of the algorithm selection problem, where the goal is to choose the most suitable solver for a given problem instance. Traditional approaches rely on supervised learning models, while this work explores whether general-purpose LLMs, without task-specific training, can perform this task by interpreting structural descriptions of constraint models. The experimental evaluation is conducted on benchmark instances from the 2025 MiniZinc Challenge. Several input representations are investigated, including raw MiniZinc scripts, structured numerical features extracted with mzn2feat, and natural-language descriptions derived from models. Additionally, a deterministic transformation tool, fzn2nl, is introduced to generate compact natural-language abstractions of FlatZinc programs.