- author = "Dimitrios Stamatios Bouras and Justyna Petke and Sergey Mechtaev",
- title = "{LLM}-Assisted Crossover in Genetic Improvement of Software",
- booktitle = "14th International Workshop on Genetic Improvement @ICSE 2025",
- year = "2025",
- editor = "Aymeric Blot and Vesna Nowack and Penn {Faulkner Rainford} and Oliver Krauss",
- pages = "19--26",
- address = "Ottawa",
- month = "27 " # apr,
- note = "Best presentation",
- keywords = "genetic algorithms, genetic programming, Genetic Improvement, MAGPIE, Large Language Models, LLM, ANN, Linux runtime, Weka, GPT4.mini",
- isbn13 = "979-8-3315-0193-8",
-
URL = "
https://gpbib.cs.ucl.ac.uk/gi2025/bouras_2025_GI.pdf",
-
URL = "
https://geneticimprovementofsoftware.com/events/icse2025#accepted-papers",
-
DOI = "
doi:10.1109/GI66624.2025.00012",
-
video_url = "https://www.youtube.com/watch?v=rCc4AvX227I&list=PLI8fiFpB7BoL5jUkvndWvSEywsu8dMdIo&index=6",
-
video_url = "https://gpbib.cs.ucl.ac.uk/gi2025/gi_2025_slides/bouras_icse2025.mp4",
-
slide_url = "https://gpbib.cs.ucl.ac.uk/gi2025/gi_2025_slides/bouras_icse2025_slides.pdf",
-
code_url = "https://github.com/SOLAR-group/LLM_Assisted_Crossover",
- size = "8 pages",
- abstract = "We evaluated against five traditional crossover methods across seven benchmarks, measuring performance on four key metrics: average ranking, best variant execution time, efficiency in reaching performance milestones, and viable variant count. Results show that LLM-assisted crossover achieved an average ranking of 2.27 (on a scale where 1 is best and 6 is worst), making it the top-performing method across benchmarks based on the quality of the optimal variants produced. The LLM-based approach also improved the fitness (execution time) by an average of 8.5 percent over the best variant produced by the traditional methods. In terms of efficiency, the LLM-assisted crossover required on average 25.6 percent fewer variants to reach 25 percent, 50 percent, 75 percent, and 100 percent of the final performance improvement, compared to the traditional methods. Additionally, the LLM-assisted crossover produced 4.8 percent more viable variants across scenarios, including both source code modification and parameter tuning cases. These findings suggest that LLMs can significantly enhance genetic programming by guiding the crossover process toward more effective and viable solutions, providing motivation for further research in LLM-assisted evolutionary algorithms.",
-
notes = "Video chair Carol Hanna.
Q: Multi-parent crossover? A: up to 5 parents, here 2 parents, next number of parents limited by LLM token window.
Q: LLM versus local search? A: next LLM with local search
Q: Slide 5, LLM access to documentation (usually XO no such data). A very long prompt? A: Original program plus, Parent 1 and 2 as only edits made (so small), so in practise no problem with length of prompt. Q: What of very big systems. Limit prompt to context around point of interest? A: Limit prompt to window in source code before and after edit being made.
Q: A: extension imminent at UCL
GI @ ICSE 2025, part of \cite{blot:2025:GI}",
