Reproduce “OptiMUS: Optimization Modeling Using MIP Solvers and large language models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ali AhmadiTeshnizi, Wenzhi Gao, Madeleine Udell

Researchers at Stanford University developed OptiMUS, an LLM-based agent that automates the formulation and solving of Mixed-Integer Linear Programming problems from natural language descriptions. The system combines structured inputs, iterative debugging, automated testing, and problem rephrasing, significantly improving success rates, with GPT-4 solving nearly twice as many problems as a basic LLM prompting strategy.

Key insights

  • Structured input formats (SNOP) effectively mitigate LLM context window limitations and guide models towards more accurate problem interpretations.
  • Iterative debugging and automated testing capabilities empower LLMs to self-correct generated code, leading to more robust and reliable solutions.
  • Problem rephrasing (augmentation) is a highly effective strategy, especially for less capable LLMs, significantly boosting success rates by allowing multiple attempts at interpretation and formulation.
  • LLMs can act as intelligent orchestrators, translating natural language into formal mathematical models and executable code for powerful external optimization solvers, rather than solving problems directly.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.