The problem the researchers set out to solve is deceptively simple to state. Large language models are remarkably good at reading a paragraph about delivery routes, production schedules, or network flows and producing plausible mathematical notation. But plausible is not correct. When a model hallucinates an objective coefficient, misapplies a constraint, or silently drops a variable definition, the error flows downstream into the generated code, and the solver returns a confidently wrong answer. Traditional single-pass pipelines offer no checkpoint where such mistakes can be caught, and existing evaluation metrics often judge only whether the final code executes, ignoring whether the underlying mathematical structure is sound.

Chiang and Fan’s answer is a decoupled, dual-team architecture. The first team, the logical abstraction team, formalizes the natural language problem into a mathematical program. It is staffed by a modeling coordinator, specialized processors — a terminology processor that resolves domain-specific vocabulary and a parameter processor that surfaces the numerical data embedded in the problem text — and, most importantly, a model auditor. The second team, the physical implementation team, transcribes the audited model into solver-ready code, with an implementation expert that classifies and repairs execution faults and evaluators that check both syntax and semantics of the generated program. The audited mathematical model is the sole interface between the two teams, which means implementation errors cannot contaminate the logical formulation and vice versa.

The auditing protocol is the framework’s technical core. The model auditor jointly checks semantic fidelity — whether the mathematical model actually says what the problem text says — and the consistency of variable definitions. On the code side, an execution-based syntax evaluator runs the generated program to catch crashes and illegal constructs, while a semantic evaluator probes whether the code’s behavior matches the intended model. Together these checks form a pipeline of verification stages that a formulation must survive before it is considered finished.

Perhaps the most novel contribution is a structural quality metric built on Graph Edit Distance. Rather than relying on manual annotation of problem texts, the researchers represent each mathematical model as a graph whose nodes and edges encode variables, constraints, and objective terms, and then measure the minimum edit cost separating a generated model from a reference formulation. Because exact Graph Edit Distance computation is NP-hard, the protocol enforces a fixed 180-second search budget per pair. A search that terminates returns the exact distance; one that exhausts the budget returns the cheapest complete edit path found — a certified upper bound on the true distance. The researchers verified through systematic probes at four times the budget that these bounds introduce negligible slack and do not alter the ordering of competing methods. A method that fails to produce a parseable model at all is charged a normalized distance of 1.0, which the authors show is not an arbitrary penalty but simply the value the metric already assigns to the empty graph — and, if anything, a lenient one, since some delivered formalizations on the hardest dataset measure above 1.28.

Experiments were conducted on three public benchmarks for (mixed-integer) linear programming: NL4Opt, with 231 instances; ComplexOR, with 18 complex operations research problems; and MAMO, with 115 mathematical modeling problems. Across all three, the dual-team framework reduced implementation failures and improved both the accuracy of the mathematical formulation and the correctness of the final solver code, measured by objective values. Statistical testing with exact McNemar tests and Holm correction showed the framework’s accuracy significantly higher than all five baselines on ComplexOR and MAMO, and higher than four of five on NL4Opt.

The computational economics are equally notable. Multi-agent systems are inherently more expensive than single-call pipelines, but the framework issues a roughly constant number of LLM calls — about nine per instance — regardless of problem difficulty. The comparable multi-agent baseline, Chain-of-Experts, scales from 8 calls on NL4Opt to 21 on MAMO. On both complex datasets, the dual-team method consumed fewer calls, fewer tokens, and less wall-clock time than Chain-of-Experts while attaining higher accuracy, demonstrating that the reported gains do not stem from an inflated call budget.

Two detailed case studies in the paper’s appendix illustrate exactly how the auditor rescues formulations that would otherwise fail silently. In one, a budget-constrained cell-tower coverage problem, the unaudited pipeline hallucinated an objective coefficient of 250 where the problem text supports only 100, and added vacuous coverage constraints that binary variables trivially satisfied. The auditor corrected the coefficient and dropped the spurious rows, restoring the verified optimum of 200. In a second case, a maximum-flow problem, the unaudited pipeline imposed flow-conservation equality on every node — including the source and sink, whose imbalance is precisely the quantity being maximized. This category error drove all flow to zero and degenerated the model, yet the over-constrained program still executed without complaint. The auditor, recognizing that conservation applies only to intermediate nodes, recovered the correct maximum flow of 11. Neither error would have been caught by execution alone.

The framework’s scope is deliberately bounded. The auditor performs value, sense, type, bound, and spurious-element corrections rather than inventing new modeling content, keeping the system within a verifiable envelope. The authors also acknowledge limitations candidly: on NL4Opt, the CAFA baseline achieved higher accuracy by a small margin, and on the smallest dataset, problem-level statistical significance was reached against only one baseline. The processed and cleaned dataset generated during the study, including corrected annotations and formatting adjustments, is publicly available to support replication.

What emerges is a template for how large language models might be organized for tasks where correctness is non-negotiable. Rather than asking a single model to do everything — read, formalize, code, and verify — the dual-team design assigns each function to a dedicated agent, inserts an audited artifact as the sole handoff between abstraction and implementation, and measures structural fidelity with a metric that does not depend on manual annotation. As optimization modeling from natural language matures from research demo toward practical tool, the lesson of this work is that the architecture of cooperation matters as much as the capability of the underlying models.

Subject of Research: Automated optimization modeling and auditing via dual-team multi-LLM-agent cooperative framework

Article Title: Automated optimization modeling and auditing via dual-team multi-LLM-agent cooperative framework

Article References: Chiang, Y., & Fan, W. (2026). Automated optimization modeling and auditing via dual-team multi-LLM-agent cooperative framework. Applied Intelligence, 56(16), Article 498. https://doi.org/10.1007/s10489-026-07536-z

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07536-z

Keywords: large language models, multi-agent systems, optimization modeling, automated auditing, mathematical programming, linear programming, mixed-integer programming, Graph Edit Distance, NL4Opt, MAMO, ComplexOR, solver code generation