OPTAgent: LLM Agents for Solver Ecosystems
The optimization toolbox keeps growing: more solver methods, more configurations, more complementary procedures. Each is usually developed and benchmarked on its own, which leaves practitioners with a hard question. Which method should be used on this instance, and with what share of a limited, shared compute budget?
OPTAgent (joint with Hanyang Jiang and Miao Li as co-first authors, and with Danlin Li, Sikai Cheng, Pascal Van Hentenryck, and Russell Bent) is an agentic policy-synthesis framework. During offline discovery, a large language model proposes and searches a space of compact decision-tree policies. Each policy selects and configures optimization methods and allocates the computational budget across a whole collection of problems.
Key ideas:
- Shift the unit of design from a standalone solver to a deployment-aware ecosystem of solvers
- Use an LLM as a search engine over interpretable decision-tree policies rather than as a solver itself
- Treat compute availability as a first-class part of method selection and configuration