Choosing an efficient formulation of an optimization model is vital, in many cases for producing valuable solutions to optimization models yet requires expert knowledge that is scarce in practice. We investigate whether large language models (LLMs) can fill this gap by autonomously proposing more computationally efficient formulations of an existing model. We propose an approach that prompts an LLM with the source code of a baseline GurobiPy implementation and a natural language problem description to generate several efficient candidate reformulations.

Each candidate is automatically validated for correctness and benchmarked against the baseline across multiple problem instances. To carry out this assessment, and to support similar evaluations by others, we introduce a benchmark dataset of 10 practical mixed-integer optimization problems, each with an inefficient baseline formulation, spanning domains including logistics, healthcare, and resource allocation. Across these problems, the best LLM-proposed reformulation reduces the solver’s computational effort by a median of 93% relative to the baseline.

This is joint research with Nathan Ferchtandiker (University of Amsterdam), Madeleine Udell (Stanford University), and Segev Wasserkrug (IBM).

This webinar is part of the Gurobi Days Digital Academic webinar series focused around the theme of AI Meets Optimization: New Ideas Transforming Research and Education. You can register for our second event of this series here: https://www.gurobi.com/resources/webinar-events/ai-future-optimization-gdda2026