Operations Research 2025
Abstract Submission

2271. Reinforcement Learning for Teaching LLMs to Derive Linear Programs

Invited abstract in session TA-12: AI for Optimization Modeling, stream Artificial Intelligence, Machine Learning and Optimization.

Thursday, 8:45-10:15
Room: H10

Authors (first author is the speaker)

1. Florian Roland Breda
University of Siegen
2. Ulf Lorenz
Chair of Technology Management, Universitaet Siegen

Abstract

Large Language Models (LLMs) have the potential to assist humans in solving complex problems objectively and efficiently, even when users lack formal knowledge of mathematical optimization. However, LLMs do not yet possess an inherent understanding of linear programs (LPs). Enabling LLMs to autonomously derive LPs from textual problem descriptions requires a structured learning approach. A reinforcement learning environment provides rewards based on the quality of generated LP formulations, guiding the learning process. The quality is assessed based on factors such as solvability, solution time, and the clarity of the formulation’s explanation. Small, specialized learning environments could be particularly useful for tackling specific problem domains.

Keywords

Status: accepted


Back to the list of papers