From drug synthesis planning and production scheduling to facility location problem, combinatorial optimization (CO) tackles decision-making over discrete solution spaces under constraints. Despite decades of progress, high-performing CO solutions typically rely on expert-crafted formulations, specialized solvers, and time-consuming parameter tuning — costs that limit adaptability in complex, large-scale, and dynamic settings.
A new review by researchers from Tsinghua University, Beijing University of Technology, and the National University of Defense Technology synthesis how LLMs are being integrated into CO pipelines to reduce expert burden and improve adaptability. The review, titled "Advancec in LLM-Assisted Combinatorial Optimization," is published in "Complex System Modeling and Simulation" [2026 Volume 6, Number 3]. Leveraging their ability to understand natural language specifications, generate executable code, and interact with external tools, LLMs are increasingly used as assistant components across the solver workflow.
The authors propose a methodology-oriented taxonomy that groups recent work into three mainstream paradigms:
- LLM-assisted modeling and solver collaboration, where an LM translate natural-language requirements into a formal optimization model or solver-ready code, and an external solver provides executable feedback (such as syntax errors, infeasibility, or objective quality) to guide iterative repair. This paradigm includes both inference-orchestration-based closed-loop refinement and parameter-adaptation-based approaches that fine-tune models to reduce trail-and-error at inference time.
- LLM-assisted heuristics, where LLMs help invent heuristic components, control search bu selecting operators or parameters, or execute as a module that proposes candidate updates — often with an outer loop that evaluates generated heuristics on benchmark instances and retains strong performers.
- LLM-based end-to-end solvers, which attempt to map an instance directly to a solution through learned inference, either by directly decoding solutions or by generating intermediate reasoning traces, aiming for fast amortized solving after training.
"Combinatorial optimization is widely used, but strong performance often still requires substantial expert intervention in modeling, algorithm design, and tuning," said Ling Wang, professor in the Department of Automation at Tsinghua University and corresponding author of the review. "Our goal was to clarify where LLMs can enter the workflow, how they interact with solvers, and what trade-offs each integration pattern brings feasibility, scalability, and reliability."
The review also highlights a fast-growing application landscape, including planning and scheduling, wireless network optimization, and engineering design optimization — areas where LLMs often work best as assistants that translate intent into formal structures, generate candidate strategies, and coordinate existing solvers rather than replacing them outright.
Looking ahead, the authors emphasize three practical bottlenecks: (1) missing pr incorrect constraints in automatic modeling; (2) high evaluation cost in heuristic-generation outer loops, and (3) limited reliability mechanisms such as uncertainty calibration and scalability to industrail-size instances. They argue that progress will likely depend on architectures that explicitly separate responsibilities: LLMs for high-level interpretation and strategy generation, and classical solvers or specialized controllers for fast, verifiable optimization.
The research team expects the review to offer a unified perspective on how LLMs can be effectively embedded into combinatorial optimization, and to guide future work toward more reliable and scalable integration. "LLM-assisted combinatorial optimization is opening a promising path toward more flexible, human-aligned decision-making, and we believe this interdisciplinary direction will continue to grow in both depth and practical impact," said Ling Wang.
Other contributors include Chuyue Tian, Rui Li from Tsinghua University, Jingfang Chen from Beijing University of Technology, and Jianmai Shi from National University of Defense Technology.
This work was supported by the National Natural Science Foundation of China (Nos. U24A20273, 62273193, and 62403272).
DOI Link:
https://doi.org/10.23919/CSMS.2026.0007
