The research, led by Huatian Gong of Nanyang Technological University together with Shuaian Wang, Dongping Song, Jiuh-Biing Sheu and Ran Yan, begins from an observation that will resonate with anyone who has tried to use an AI assistant for serious programming. If you simply ask a large language model to produce a complete heuristic for a combinatorial optimization problem in a single pass, the result usually fails. The model may misunderstand the input format, mishandle edge cases, or produce code that cannot run at all. The failure is not primarily a shortage of intelligence but a shortage of structure: the model is being asked to solve the mathematical problem and the software engineering problem simultaneously, with no verified contract separating the two.

The team’s answer is a framework called LACE, short for LLM-driven Algorithm Construction via Complementary Evolution. Its first ingredient is what the authors call the I-O-T-H interface, a formal problem contract with four components: an input schema, an output schema, a tool library, and a heuristic portfolio. The input and output schemas pin down exactly what data the heuristic receives and what it must return, eliminating an entire class of implementation errors. The tool library supplies verified building blocks, common routines such as evaluation functions and local search moves, so the model does not have to reinvent and re-verify basic machinery every time. The heuristic portfolio, finally, is the heart of the system: rather than betting everything on one algorithm, LACE maintains a collection of specialist heuristics, each of which may excel on a different subset of problem instances. With this contract in place, the language model’s capacity is redirected away from low-level plumbing and toward the high-level algorithmic reasoning where it genuinely adds value.

The second ingredient is the evolutionary engine. Under a strict runtime budget for each problem instance, LACE iteratively generates candidate heuristics, tests them, and selects a portfolio of specialists whose strengths cover heterogeneous cases. The word complementary is doing real work here. A single heuristic that performs well on average can be worse than a team of heuristics that each dominate on particular instance types, provided the system can decide which specialist to deploy. LACE’s complementary selection mechanism explicitly optimizes for coverage of the instance distribution rather than for a single champion, an approach that echoes the hyper-heuristics tradition in operations research but replaces human-designed selection rules with an automated, model-driven search. The framework uses four designer agents in its first stage and seven evolution operators in its second, all of which the authors have released openly.

The evaluation is unusually thorough. The researchers tested LACE on 36 classical problems drawn from CO-Bench, a benchmark suite designed to measure how well language model agents can search for algorithms on combinatorial optimization tasks. These problems span the canonical territory of the field, including routing, scheduling, packing and assignment variants that have accumulated decades of human algorithmic effort. On this suite, LACE achieved an average score of 0.945. The strongest existing LLM-based method reached 0.870, while direct prompting of a language model without any supporting framework managed only 0.571. The gap between the framework and bare prompting is the study’s central message: the gain comes from the architecture, not from a smarter model. The same underlying language model, given the right scaffolding, performs dramatically better than the same model asked to improvise.

Even more striking is the result on generalization. The team constructed four structurally new optimization problems that the models had never seen during development, the kind of novel variants that arise constantly in industry when a business’s constraints do not match any textbook problem. On these four problems, LACE reached scores between 0.97 and 0.99, while five existing LLM-based baselines failed to produce any feasible algorithm at all. This is the difference between a system that has memorized solutions to famous problems and one that can genuinely engineer an algorithm for an unfamiliar contract. For logistics and manufacturing, where bespoke constraints are the norm rather than the exception, that distinction is the whole ballgame.

The authors also examined robustness across different frontier language model backbones and found that LACE’s performance holds, with varying cost-efficiency trade-offs depending on which model powers the framework. This matters because it suggests the contribution is durable: as models improve or change, the interface-and-evolution architecture should continue to extract value from whatever model sits underneath. The ablation studies reinforce the point that both major components are essential. Removing the tool library degrades performance, and so does removing the complementary portfolio, confirming that verified building blocks and specialist diversity contribute independently to the framework’s success.

The transparency of the project is itself noteworthy. All code, including the designer agents, the evolution operators, the complementary-selection solver and the scripts reproducing every figure, is available under an MIT licence on GitHub and Zenodo, along with the instance sets for all 40 problems, the evolved heuristic portfolios and the per-instance results. A Colab notebook allows anyone to reproduce the reported results without local installation, and an interactive supplementary webpage presents all per-instance outcomes. In a field where claims of automated scientific discovery sometimes rest on opaque pipelines, this level of openness invites scrutiny and reuse in equal measure. The work was supported by Singapore’s Agency for Science, Technology and Research, the Japan Science and Technology Agency, the Ministry of Education of Singapore and the UK Engineering and Physical Sciences Research Council.

The broader significance extends beyond the benchmark numbers. Algorithm design has long been a bottleneck at the intersection of operations research, computer science and industry: the mathematics of a problem may be well understood, yet translating that understanding into a fast, reliable solver remains skilled manual labor. The LACE results indicate that a well-structured division of labor between human-designed contracts and machine-generated heuristics can compress that labor dramatically. The verified interface acts as the human contribution, encoding what a correct solution looks like, while the evolutionary search acts as the machine contribution, exploring the space of algorithmic strategies far faster than a human team could. It is a template that seems likely to spread to neighboring domains, from constraint programming to simulation optimization, wherever a problem can be specified precisely enough to form a contract.

There are, of course, caveats worth keeping in view. The benchmark scores measure performance within defined runtime budgets on defined instance distributions, and real-world deployments will bring messier data, shifting constraints and integration challenges that no offline benchmark fully captures. The framework still depends on capable language models, with all their costs and failure modes, and the complementary selection adds computational overhead of its own. Yet the direction of travel is clear and the evidence is strong. What the study demonstrates is that the discovery of effective algorithms for hard combinatorial problems, long considered a craft reserved for a small community of experts, can be automated to a substantial degree. If the pattern holds as models and frameworks improve, the heuristics quietly running the world’s supply chains, hospitals and grids may increasingly be designed not by a lone expert over months, but by an evolving portfolio of machine-discovered specialists in days.

Subject of Research: Automated design of complementary heuristics for combinatorial optimization using large language models

Article Title: Large language models discover complementary heuristics for combinatorial optimization

Article References: Gong, H., Wang, S., Song, D., Sheu, J.-B., & Yan, R. (2026). Large language models discover complementary heuristics for combinatorial optimization. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01307-8

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01307-8

Keywords: large language models, combinatorial optimization, heuristics, algorithm design, LACE framework, CO-Bench, evolutionary computation, hyper-heuristics, Nature Machine Intelligence, automated scientific discovery, operations research, artificial intelligence