Property-Based Fitness from Examples: Guiding GP with LLM-Discovered Invariants

Created by W.Langdon from gp-bibliography.bib Revision:1.9194

@InProceedings{Garrow:2026:UKCI,

• author = "Fraser Garrow and Michael A. Lones and Robert Stewart",
• title = "Property-Based Fitness from Examples: Guiding {GP} with {LLM}-Discovered Invariants",
• booktitle = "25th UK Workshop on Computational Intelligence (UKCI 2026)",
• year = "2026",
• address = "Coventry University",
• month = "9--11 " # sep,
• keywords = "genetic algorithms, genetic programming, grammatical evolution, GP-LLM, AI, ANN, program synthesis, large language models, LLM, OpenAI GPT-5.0, property-based testing, program invariants, Abstraction and Reasoning Corpus (2D images), ARC, BBE (Super Anagrams, Vector Summed, and Smallest), XAI, Python",
• size = "14 pages",
• abstract = "Program synthesis via Genetic Programming (GP) traditionally relies on Programming by Example (PBE), which often struggles to capture a target function underlying logic, leaving models susceptible to overfitting. Semantically-aware fitness functions improve generalisation but typically require manual expert specification. We propose a hybrid framework that automates this process by repurposing Large Language Models (LLMs) as analytical tools. Given only input-output examples, an LLM infers general program invariants, which are then translated into executable property-based tests and integrated into a GP fitness function. Evaluated on standard synthesis benchmarks and the Abstraction and Reasoning Corpus (ARC), our findings demonstrate that LLM-discovered properties act as effective soft regularisers, mitigating the standard PBE overfitting problem. While LLM-inferred properties can themselves contain semantic biases, decaying their influence over the evolutionary run enables them to provide early-search scaffolding without permanently trapping the population. This approach steers evolution toward more robust solutions, highlighting synergy between the semantic reasoning of LLMs and the exploratory search of GP.",
• notes = "Mike Ernst Daikon 'promising synergy between the exploratory power of GP and the semantic reasoning capacity of LLMs.'

  'executable PBTs.'

  \cite{Helmuth:2015:GECCO} but see also \cite{Helmuth:2022:GPEM} etc

  weighted fitness = changing linear combination of fitness on test cases and fitness on LLM's inferred properties

  'Decaying the property weight allows the system to exploit initial semantic guidance while retaining the freedom to discard LLM biases later in the search.'

  To be published after the workshop

  https://ukci2026.coventry.ac.uk/",

}

Genetic Programming entries for Fraser Garrow Michael A Lones Robert Stewart

Citations