Two smart agents solve optimization dilemmas with processor cores: the SailSearch framework resets computing budgets in the region
Listen to this article
Read by Anchor
A research team that includes Dahrie Ferma, Kanthi Sarbatur, Venisius Lima, Dzhong Van and Giant Kalagnanam, via the Archive platform, presented a framework called SailSearch that combines human guidance with two language-model agents to automate algorithmic search in combinatorial optimisation problems. This development opens new possibilities for engineering and technical teams in the Gulf region and Egypt to solve complex mathematical challenges in supply chains and transport networks without needing costly server farms, after the system demonstrated its ability to solve historically intractable problems using modest central processors.
Unprecedented solutions on just two cores.The paper focused on testing the framework on two core problems, the first being the Market Split problem. Using the two agents, the system developed a stochastic algorithm that searches across multiple graph bases and allocates computational effort precisely between enumeration and continuous reduction of the bases. Across ten random seeds per case, the algorithm recorded the first documented solutions that rely exclusively on central processors for eighteen problems that had remained unsolved out of thirty-two in the standard QOBLIB library. It also solved between one hundred and thirty and one hundred and thirty-three of the fifty-six benchmark problems per random seed, within a 24-hour deadline and using only two CPU cores.
Solution software shows shortcomings on graphs.The second problem was the Maximum Independent Set (MIS) challenge. The researchers built a generator that encodes XOR constraints as graphs with mathematically proven upper bounds, and tested seven graph instances ranging from 1,752 to 3,996 vertices. Results showed that the best solutions returned by the tested mathematical solvers after eight-hour runs were 5 to 15 vertices short of the proven optimal, revealing real accuracy gaps in conventional commercial and academic solvers when handling large networks.
The study demonstrates that involving language-model agents under human supervision goes beyond traditional code generation to explore new mathematical structures that reduce computational waste. Instead of resorting to expensive graphics-processing units, the innovative algorithms re-evaluate ordinary CPUs through a carefully balanced mathematical partitioning of workloads.
What this means for the region
The achievement represents a tangible shift for infrastructure and technology managers in logistics, ports and telecommunications in Saudi Arabia, the United Arab Emirates and Egypt. Scheduling shipments, fleet routing and frequency allocation all depend structurally on combinatorial optimisation, and the study’s results suggest that capital spending on accelerating cloud computing could be cut, as intelligent algorithms enable complex problem solving on ordinary local servers with the fewest possible CPU cores.
The findings also place technical audit teams in banks and regional network firms before a need to reassess off-the-shelf solution tools. Demonstrating that current solvers fall short of optimal results on large graphs by up to 15 vertices means that systems presently used for distribution-network management or risk assessment may waste operational efficiency without the organisations’ knowledge, prompting the adoption of mathematically validated benchmark tests to evaluate purchased solutions.
Practical value today shifts from manually tweaking equations or buying closed-source software to enabling engineers to direct AI agents to explore untrodden algorithmic paths and design self-verification tests, giving regional platforms greater flexibility and lower operating costs.