OPTAgent: From Standalone Solvers to Deployment-Aware Solver Ecosystems
Published in NeurIPS 2026 Workshop on ML×OR, 2026
Hanyang Jiang*, Changkun Guan*, Miao Li*, Danlin Li, Sikai Cheng, Pascal Van Hentenryck, Russell Bent. "OPTAgent: From Standalone Solvers to Deployment-Aware Solver Ecosystems." NeurIPS 2026 Second Workshop on ML×OR: Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making, 2026. (*Equal contribution)
Decades of numerical optimization research have produced a large toolbox of solver methods, configurations, and complementary procedures, mostly developed and evaluated one algorithm at a time within fixed problem settings. That makes it increasingly hard to know which method to use on a particular problem instance. Available compute adds another dimension, especially when many problems must be solved in limited time on shared resources.
OPTAgent is an agentic policy-synthesis framework that uses large language models to build method-selection and configuration policies from a toolbox of optimization solvers. During offline discovery, an LLM proposes and searches a space of compact decision-tree policies that allocate a computational budget to methods and their configurations across collections of problems.
To be presented as a poster at the Second Workshop on ML×OR at NeurIPS 2026.