Researchers have developed an automated approach to designing feature extractors for constraint satisfaction problems. This method uses large language models in a loop to generate executable code that captures problem structure. The synthesized extractors outperform hand-curated features and state-of-the-art transformer-based variants, yielding algorithm selectors with improved test-set accuracy. The approach remains interpretable and can be applied to new problem classes. --- Why it matters: This matters because it addresses the bottleneck of manually designing feature extractors for complex problems, enabling researchers to focus on other aspects of algorithm selection. It also demonstrates the potential of large language models in synthesizing domain-specific code. Source: https://arxiv.org/abs/2608.17170