NEMO: Execution-Aware Optimization Modeling via Autonomous Coding Agents

NEMO is an AI system that converts natural language descriptions of optimization problems into executable mathematical code using autonomous coding agents. The approach achieves state-of-the-art results on optimization benchmarks by treating code execution as a first-class constraint, ensuring generated solutions are functional by design rather than relying on specialized language models that often produce broken code.

NEMO addresses a fundamental limitation in AI-assisted optimization: the gap between language model outputs and executable implementations. Traditional approaches using specialized LLMs frequently generate syntactically invalid or non-functional code, requiring manual debugging and repair. NEMO inverts this problem by using sandboxed autonomous coding agents where execution becomes a validation mechanism rather than an afterthought, guaranteeing code functionality by construction.

The system's architecture incorporates several sophisticated innovations that reflect maturing agentic AI design patterns. Asymmetric validation loops between independently generated optimizer and simulator implementations create mutual verification mechanisms. External memory systems enable experience reuse across problem instances, improving performance on related tasks. Robustness enhancements through minimum Bayes risk decoding and self-consistency further reduce failure modes. These architectural choices represent an evolution beyond single-pass generation toward iterative refinement grounded in executable feedback.

For the optimization and operations research communities, NEMO demonstrates that execution-aware design principles substantially outperform traditional language model approaches on established benchmarks. This validates a broader trend where constraining AI systems to produce verifiable artifacts yields more reliable automation. The achievement across nine diverse optimization benchmarks suggests the approach generalizes beyond narrow domains, potentially enabling broader automation of mathematical modeling workflows that previously required human expertise.

The technical contributions point toward future AI systems where sandboxed execution environments become standard infrastructure for ensuring code reliability. Organizations relying on automated optimization could accelerate deployment timelines by reducing validation overhead, though production adoption will depend on how the system handles domain-specific constraints and real-world problem complexity.

• →NEMO achieves state-of-the-art optimization modeling by treating code execution as a validation constraint rather than a post-hoc requirement.
• →Autonomous coding agents with sandboxed execution guarantee syntactically valid implementations, eliminating a major failure mode of specialized LLM approaches.
• →Asymmetric validation loops and external memory mechanisms enable robust coordination between independent optimizer and simulator components.
• →Performance gains across nine established benchmarks suggest the execution-aware architecture generalizes across diverse optimization problem classes.
• →The approach represents a shift toward grounding AI code generation in verifiable execution environments rather than relying solely on language model outputs.