TL;DR: Researchers published a framework called LACE on October 1, 2026, that lets large language models design teams of complementary heuristics for combinatorial optimization, scoring 0.945 across 36 benchmark problems.

A published framework for automated heuristic design

A study titled Large language models discover complementary heuristics for combinatorial optimization was published on October 1, 2026, in Nature Machine Intelligence, according to the Nature Machine Intelligence article page and the journal's research-articles listing.

The authors listed for the study are Huatian Gong, Shuaian Wang, Dongping Song, Jiuh-Biing Sheu and Ran Yan, according to the Nature Machine Intelligence article page and Crossmark record for DOI 10.1038/s42256-026-01307-8.

The study introduces LLM-driven Algorithm Construction via Complementary Evolution, abbreviated LACE, according to the Nature Machine Intelligence article page.

The Nature Machine Intelligence research-articles listing describes LACE as a large language model-based framework that designs optimization algorithms by building a verified problem contract and then evolving a portfolio of complementary heuristics.

Coverage of the study in Scienmag on October 1, 2026, describes the central result as automation of heuristic design to a substantial degree, with language models discovering families of heuristics that work together under strict time limits.

The article history recorded by Crossmark shows the manuscript was received on March 8, 2026, accepted on August 25, 2026, and first published online on October 1, 2026.

The authors declare no competing interests, according to the Crossmark record.

Why combinatorial optimization needs new design methods

Combinatorial optimization underpins decision-making across science, engineering and industry, especially manufacturing, logistics, healthcare operations and energy management, according to the Nature Machine Intelligence article page.

Combinatorial optimization is described by Scienmag as the invisible machinery involved when a factory schedules production lines, a shipping company routes vessels, a hospital allocates operating rooms, or a power grid balances fluctuating loads.

Those problems share a finite set of choices, a scoring function, and a combinatorial explosion that makes checking every option impossible, according to Scienmag.

For decades, tackling a new variant has relied on a human expert spending months designing a heuristic, a rule of thumb that trades guaranteed optimality for speed and practical quality, according to Scienmag.

Designing effective heuristics for each new problem variant requires deep domain knowledge and months of expert iteration, according to the Nature Machine Intelligence article page.

Directly prompting large language models can generate executable code, but single-pass synthesis of complete heuristics routinely fails on new combinatorial optimization problems, according to the Nature Machine Intelligence article page.

Scienmag likewise reports that asking a large language model to produce a complete heuristic in a single pass usually fails, with possible misunderstanding of input format, mishandling of edge cases, or code that cannot run at all.

The I-O-T-H interface and verified problem contract

LACE decomposes algorithm design into an input schema, output schema, tool library and heuristic portfolio, termed the I-O-T-H interface, according to the Nature Machine Intelligence article page.

Scienmag describes the same four components as an input schema, an output schema, a tool library, and a heuristic portfolio, forming a formal problem contract.

The interface specifies a verified problem contract under which the language model operates, according to the Nature Machine Intelligence article page.

The contract directs model capacity to high-level algorithmic reasoning rather than problem-specific implementation, according to the Nature Machine Intelligence article page.

Scienmag reports that the input and output schemas pin down exactly what data a heuristic receives and what it must return, eliminating an entire class of implementation errors.

The tool library supplies verified building blocks, including common routines such as evaluation functions and local search moves, so the model does not have to reinvent basic machinery every time, according to Scienmag.

The heuristic portfolio maintains a collection of specialist heuristics, each of which may excel on a different subset of problem instances, according to Scienmag.

Scienmag reports that, with the contract in place, language-model capacity is redirected away from low-level plumbing and toward high-level algorithmic reasoning.

Complementary evolution under strict runtime budgets

LACE refines the portfolio through time-constrained complementary evolution, according to the Nature Machine Intelligence article page.

Complementary evolution iteratively generates and selects specialist heuristics capable of solving heterogeneous instances within strict runtime budgets, according to the Nature Machine Intelligence article page.

Scienmag describes the evolutionary engine as iteratively generating candidate heuristics, testing them, and selecting a portfolio of specialists whose strengths cover heterogeneous cases under a strict runtime budget for each problem instance.

The word complementary reflects that a team of heuristics that each dominate on particular instance types can outperform a single heuristic that performs well on average, provided the system can decide which specialist to deploy, according to Scienmag.

LACE's complementary-selection mechanism explicitly optimizes for coverage of the instance distribution rather than for a single champion, according to Scienmag.

Scienmag places that approach in the hyper-heuristics tradition in operations research, while noting that LACE replaces human-designed selection rules with automated, model-driven search.

The study's code release includes four stage-one designer agents, seven stage-two evolution operators, and a complementary-selection solver, according to the Nature Machine Intelligence article page.

Results across 36 classical CO-Bench problems

The researchers tested LACE on 36 classical problems drawn from CO-Bench, according to the Nature Machine Intelligence article page.

CO-Bench is described by Scienmag as a benchmark suite designed to measure how well language-model agents can search for algorithms on combinatorial optimization tasks.

The 36 classical benchmark problems are redistributed from CO-Bench, according to the Nature Machine Intelligence article page.

Scienmag reports that these problems span routing, scheduling, packing and assignment variants that have accumulated decades of human algorithmic effort.

Across the 36 classical CO-Bench problems, LACE attained an average score of 0.945, against 0.870 for the strongest existing large language model-based method and 0.571 for direct language-model prompting without any framework, according to the Nature Machine Intelligence article page.

Scienmag reports the same three figures: 0.945 for LACE, 0.870 for the strongest existing language-model-based method, and 0.571 for direct prompting without a supporting framework.

The study locates the gain in the framework rather than the model, according to the Nature Machine Intelligence article page.

A figure caption in the Nature Machine Intelligence article page states that LACE outperforms existing language-model-based and classical methods across the CO-Bench suite.

Generalization to four structurally new problems

On four structurally new problems, LACE reached scores of 0.97 to 0.99, while five existing language-model-based baselines failed to produce any feasible algorithm, according to the Nature Machine Intelligence article page.

A figure caption in the Nature Machine Intelligence article page states that LACE generalizes zero-shot to four new combinatorial optimization problems.

Scienmag describes the four problems as new variants intended to test whether the system can handle structure not seen during development of established methods.

The result demonstrates that effective algorithm discovery for such problems can be automated to a substantial degree, according to the Nature Machine Intelligence article page.

Scienmag summarizes the implication as language models not merely writing code on demand but discovering genuinely complementary families of heuristics that work together under strict time limits.

Tool library, portfolio, and model robustness

A figure caption in the Nature Machine Intelligence article page states that the tool library and the complementary portfolio are both essential for LACE.

Scienmag reports that the tool library provides verified routines and that the portfolio provides specialist coverage, making both parts central to the reported performance.

Another figure caption in the Nature Machine Intelligence article page states that LACE is robust across frontier language-model backbones with varying cost-efficiency trade-offs.

That robustness finding indicates the reported behavior was observed with more than one underlying model family, with differences in cost and efficiency, according to the Nature Machine Intelligence article page figure description.

The framework architecture itself is presented in a figure titled LACE framework architecture, according to the Nature Machine Intelligence article page.

Open data, code, and reproducibility materials

All data required to interpret, verify and extend the study are openly available with no access restrictions, according to the Nature Machine Intelligence article page.

The minimum dataset comprises the instance sets for all 40 problems, the evolved heuristic portfolios and the per-instance results underlying the figures for the benchmark suite, generalization tests, ablation tests and backbone tests, according to the Nature Machine Intelligence article page.

That dataset is deposited in the LACE repository at https://github.com/PJ-NTU/LACE and via Zenodo at https://doi.org/10.5281/zenodo.21886438 under the MIT licence, according to the Nature Machine Intelligence article page.

An interactive Supplementary Information webpage presenting all per-instance results is available at https://pj-ntu.github.io/LACE/lace_supplementary_information.html, according to the Nature Machine Intelligence article page.

All codes used in the study are openly available in the same LACE repository and Zenodo record under the MIT licence, according to the Nature Machine Intelligence article page.

The released code includes the LACE framework, the four stage-one designer agents, the seven stage-two evolution operators, the complementary-selection solver and the scripts that reproduce every reported figure, according to the Nature Machine Intelligence article page.

A Colab notebook that reproduces the reported results without local installation is available at https://colab.research.google.com/github/PJ-NTU/LACE/blob/main/notebooks/reproduce.ipynb, according to the Nature Machine Intelligence article page.

Funding and publication context

Funding for the research was provided by the Agency for Science, Technology and Research, the Ministry of Education - Singapore, and the Engineering and Physical Sciences Research Council, according to the Crossmark record.

The Crossmark record lists Agency for Science, Technology and Research awards M24N8c0098 and R24I6IR135, Ministry of Education - Singapore award SSRC2024-SSHRTG-008, and Engineering and Physical Sciences Research Council award EP/Y024605/1.

The research was led by Huatian Gong of Nanyang Technological University together with Shuaian Wang, Dongping Song, Jiuh-Biing Sheu and Ran Yan, according to Scienmag.

Text and data mining permission and the version of record were valid from October 1, 2026, according to the Crossmark record.

The study appears in the Nature Machine Intelligence research-articles listing dated October 1, 2026, alongside articles on medical image restoration, neutrino detectors, climate risk assessment, modular neural networks, microrobot navigation and other topics, according to that listing.

Sources referenced