AutoOR: Scalably Post-training LLMs to Autoformalize Operations Research Problems

TL;DR

AutoOR is a scalable pipeline that trains large language models to automatically formalize complex optimization problems from natural language, improving efficiency and performance across various OR benchmarks.

Contribution

The paper introduces AutoOR, a novel method combining synthetic data generation and reinforcement learning to enable LLMs to autoformalize diverse optimization problems.

Findings

AutoOR achieves state-of-the-art results on six OR benchmarks.

AutoOR matches larger models' performance with an 8B parameter model.

A curriculum RL strategy improves performance on physical dynamics problems.

Abstract

Optimization problems are central to decision-making in manufacturing, logistics, scheduling, and other industrial settings. Translating complicated descriptions of these problems into solver-ready formulations requires specialized operations research (OR) expertise, making it hard to scale. We present AutoOR, a scalable synthetic data generation and reinforcement learning pipeline that trains LLMs to autoformalize optimization problems specified in natural language across linear, mixed-integer, and non-linear categories. AutoOR generates verified training data from standard optimization forms and uses solver execution feedback as the reward signal for RL post-training. AutoOR applied to an 8B model achieves state-of-the-art or competitive results across six established OR benchmarks, matching significantly larger frontier models. For a non-linear problem class involving physical dynamics, where frontier models score near 0%, we introduce a curriculum RL strategy that bootstraps from limited initial training data to make this class tractable for post-training. We believe that methods such as AutoOR can significantly accelerate industrial decision-making with AI.