NEMO: AI System Translates Natural Language into Executable Optimization Code

NEMO is a new system developed by researchers that transforms natural-language descriptions of decision-making challenges into formal, executable mathematical optimization codes through the use of autonomous coding agents (ACAs). Unlike current methods that depend on specialized large language models (LLMs) or specific task agents, which frequently generate invalid or non-executable code, NEMO elevates ACAs to a primary abstraction similar to API interactions with LLMs. The system ensures code executability through sandboxed execution and facilitates automated validation and repair. NEMO also presents innovative coordination patterns, such as asymmetric validation loops between independently created optimizer and simulator implementations, external memory for reusing experiences, and robustness improvements through minimum Bayes risk (MBR) decoding and self-consistency. It demonstrates considerable advancements in code correctness and execution reliability across nine established optimization benchmarks.