LLM-Guided Graph Generation Enhances Structure-Based Local Improvement Methods
An arXiv preprint (2608.13333) has unveiled an automated pipeline that leverages large language models (LLMs) to create problem-independent graphs, which assist in structure-based local improvement methods for combinatorial optimization. This method, referred to as SLIM (Structure-based Local Improvement Method), tackles the issue of variable selection in large neighborhood search, traditionally dependent on random choices or specialized strategies. The pipeline is compatible with any problem formatted in MiniZinc, ensuring wide applicability. By utilizing semantic prompts for the LLM, it produces a graph generator that translates any problem instance into a standardized weighted graph, where nodes signify decision variables and edges denote constraint relationships. These graphs facilitate variable selection within SLIM, promoting a universal representation that allows consistent graph features across all problem instances. The goal is to automate variable selection strategy design, potentially enhancing efficiency in various optimization challenges. The paper was authored by researchers and is available on arXiv under the identifier 2608.13333v1.
Key facts
- arXiv:2608.13333v1 is a new preprint.
- The paper proposes an automatic pipeline for LLM-guided graph generation.
- The pipeline is problem-agnostic and works with MiniZinc format.
- LLM is prompted with semantic guidelines to produce a graph generator.
- Graphs map instances to uniform weighted graphs with nodes as decision variables and edges as constraints.
- The graphs guide the SLIM framework in variable selection.
- Weighted graphs enable shared generic representation for all problem instances.
- The approach aims to automate variable selection strategies in large neighborhood search.
Entities
Institutions
- arXiv