What is it about?
Many real-world optimization problems involve several decisions that depend on each other. For example, a company may need to decide both how much inventory to deliver to customers and how vehicles should route those deliveries. Improving one part of the problem without considering the other can lead to poor overall solutions. In this work, we introduce CoupleEvo, a method that uses large language models to automatically create and improve heuristics for several connected optimization subproblems. We compare three ways of coordinating this process: improving the heuristics one after another, alternating between them repeatedly, or improving all of them at the same time. We test these strategies on two problems involving inventory and vehicle routing, and warehouse operations with multiple robots. The results show that evolving the heuristics sequentially or iteratively is generally more reliable than evolving all of them simultaneously.
Featured Image
Photo by Aerps.com on Unsplash
Why is it important?
Large language models have already shown that they can automatically design useful heuristics for individual optimization problems. However, many practical systems consist of several strongly connected problems, and existing approaches have largely focused on designing heuristics for a single problem at a time. CoupleEvo extends this idea to settings where several heuristics must work together. Our experiments show that the way these heuristics are evolved and coordinated has a major effect on performance. Sequential and iterative approaches produced more reliable results than evolving all heuristics simultaneously, while the integrated approach was more difficult to optimize consistently. The resulting heuristics were also competitive with established problem-specific methods and, in some comparisons, outperformed classical approaches. This suggests that LLM-based heuristic design could become a useful way to develop optimization methods for complex real-world systems with interacting decisions.
Perspectives
What I find particularly interesting about this work is that it moves LLM-based heuristic design closer to the structure of real-world optimization problems. Many practical applications cannot be reduced to a single isolated decision: different optimization components influence one another, and their heuristics therefore need to work well together. One result that stood out to me was that the most integrated strategy was not automatically the most effective. Although evolving all heuristics simultaneously allows their interactions to be considered directly, it also makes the search substantially more complex. In our experiments, simpler decomposition-based strategies were more stable and often produced better results. I think this highlights an important direction for future research: not only improving the heuristics generated by LLMs, but also developing better ways to coordinate how multiple heuristics are generated and improved together.
Thomas Bömer
Karlsruher Institut fur Technologie
Read the Original
This page is a summary of: CoupleEvo: Evolving Heuristics for Coupled Optimization Problems Using Large Language Models, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3795101.3814664.
You can read the full text:
Resources
Contributors
The following have contributed to this page
