Automated operations‑research (OR) modeling requires large language models (LLMs) to translate natural‑language decision problems into correct mathematical programs. Existing approaches can improve individual formulations but usually operate on isolated tasks, retaining little reusable experience and repeatedly making similar formulation mistakes. Memory‑based methods store examples, thoughts, or insights as references, yet OR modeling needs reusable "skills" that transfer across problem narratives and guide concrete modeling decisions.
OptiSkill introduces a skill‑augmented framework that builds a hierarchical and evolving SkillBank for LLM‑based OR modeling. The SkillBank stores solver‑verified experience as reusable skills, organized into two categories:
- Global Strategies: problem‑level formulation skeletons that help the model quickly construct the overall structure.
- Step Experiences: local error‑prevention rules for specific modeling steps, such as variable naming conventions or constraint expression patterns.
The SkillBank is refined through stable batch‑level test‑time evolution: candidate skills are incorporated only after validation, ensuring reliability and expanding coverage. Experiments on eight OR modeling benchmarks show that OptiSkill markedly improves formulation accuracy across various LLM backbones, outperforms strong agentic baselines, and gains further benefits when the SkillBank is expanded. The code and data are publicly available at https://github.com/rachhhhing/OptiSkill.
Review