LLM-Driven Synthesis of Feature Extractors for Constraint Satisfaction Algorithm Selection

An innovative automated technique has been created by researchers, employing Large Language Models (LLMs) in a check-fix-verify loop to generate executable Python scripts. These scripts serve as interpretable, problem-specific feature extractors tailored for constraint satisfaction challenges. Detailed in arXiv paper 2608.17170, this method transforms a high-level MiniZinc model and its instance into code that builds a typed graph representation, calculating structural attributes like graph density, variable clustering, and constraint tightness. The evaluation encompassed three combinatorial issues—vehicle routing, car sequencing, and fixed-length error-correcting codes—utilizing a suite of five advanced solvers. The resulting extractors yielded algorithm selectors that consistently surpassed baseline methods, indicating a scalable approach to the challenges of manual feature extraction for emerging problem categories.