Large Language Models for Optimization

IOLab investigates how large language models can complement established combinatorial optimization techniques. Rather than treating an LLM as a stand-alone solver, we study its role as one component of a structured optimization system.

Our work examines LLM-based generative operators within metaheuristics, the ability of models to recall and interpret problem features, and the information encoded in their internal representations. These representations can be evaluated as surrogates for predicting instance difficulty and algorithm behavior in algorithm-selection scenarios.

The group has published a systematic review of LLMs for combinatorial optimization and experimental work on feature extraction and algorithm selection.