Evolutionary Prompting: Automating Large Language Model Queries for Machine Learning Task Identification using Grammatical Evolution

Created by W.Langdon from gp-bibliography.bib Revision:1.9212

@InProceedings{Goncalves:2026:CEC,

• author = "Carlos Goncalves and Pedro Pereira and Paulo Cortez",
• title = "Evolutionary Prompting: Automating Large Language Model Queries for Machine Learning Task Identification using Grammatical Evolution",
• booktitle = "IEEE CEC 2026",
• year = "2026",
• editor = "Nelishia Pillay and Yew-Soon Ong and Carlos Coello and Juergen Branke and Simon Lucas and Chuan-Kang Ting and Bing Xue",
• pages = "cec pap329",
• address = "Maastricht, the Netherlands",
• month = "21-26 " # jun,
• publisher = "IEEE",
• keywords = "genetic algorithms, genetic programming, Grammatical Evolution, Agentic System, Machine Learning, Prompt Engineering, LLM, AI, ANN",
• URL = "https://wcci.klinkhamergroup.com/papers/by_date.html#cec_pap329",
• URL = "https://linklings.s3.amazonaws.com/organizations/WCCI/wcci2026/submissions/stype102/gWu1C-cec_pap329s2.pdf",
• size = "7 pages",
• abstract = "We propose a novel framework for automated prompt engineering that leverages Grammar-Guided Evolutionary Algorithms to navigate the search space of Large Language Models (LLMs). As a use case, we apply the framework to the Machine Learning (ML) Goal (MLG), where a zero-shot or few-shot LLM query is used to infer the proper MLG (e.g., classification, regression, clustering) from a business need description. Our approach uses an explicit Context-Free Grammar (CFG) to enforce syntactic constraints and inject domain knowledge into the LLM prompts. The optimization engine employs a Grammatical Evolution (GE), mapping integer genotypes to valid prompt phenotypes through grammar derivation trees. We formulate the search as a Multi-Objective Optimization Problem (MOOP), seeking a Pareto-optimal set of solutions that balance two conflicting objectives: maximizing ML task accuracy and minimizing LLM inference time. To ensure robustness, the solutions are evolved using one recently proposed MLG dataset and a newly proposed adaptive sampling mechanism (to speed up the computation). Moreover, the final Pareto solutions are then explicitly evaluated on a distinct unseen MLG dataset. Overall, competitive results were obtained when comparing our GE approach with two baselines (expert prompting and an AutoML tool)",
• notes = "University of Minho, Guimaraes, Portugal",

}

Genetic Programming entries for Carlos Goncalves Pedro Pereira Paulo Cortez

Citations