Reproduce “Strategy-Aware Optimization Modeling with Reasoning LLMs” 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

Ruiqing Zhao, Fengzhi Li, Yuan Zuo, Rui Liu, Yansong Liu, Yunfei Ma, Fanyu Meng, Junlan Feng

The SAGE framework integrates explicit "Modeling Strategy" into the Large Language Model generation process for automated optimization modeling, enhancing the reliability, diversity, and solver efficiency of generated mathematical programs. It achieves an 80.3% Pass@1 accuracy on optimization benchmarks and generates formulations leading to fewer constraints and faster solve times.

Key insights

  • Explicitly incorporating high-level "Modeling Strategy" into LLM training significantly improves the correctness, diversity, and computational efficiency of generated optimization models.
  • Structured reasoning templates, which enforce an explicit strategy definition, guide LLMs to produce globally consistent and logically sound formulations.
  • Prioritizing credit assignment to early strategy decisions during reinforcement learning, through segment-weighted GRPO, is crucial for learning effective and efficient optimization modeling behaviors.

Summary by alphaXiv.

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