Year: 2026
Type: article
Abstract: Artificial intelligence for combinatorial optimization depends on high-quality, structured training data, yet transforming ambiguous natural-language descriptions into precise, executable formal models remains a major bottleneck, traditionally requiring extensive manual effort and domain expertise. We develop CPAM-LLM, an automated constraint-programming modeling framework based on large language models that converts unstructured problem specifications into validated and executable formal models. The framework integrates retrieval-augmented generation grounded in formal concept analysis for structural validation, chaos-mapping-based data augmentation to enhance cross-domain generalization, and dynamic constraint adaptation to accommodate evolving problem settings. We demonstrate its effectiveness across five scenarios, including DNA sequence design, reconfigurable energy storage systems, and unmanned vehicle logistics. The results show that the framework can automate the data-to-model pipeline while adapting reliably to changing requirements, providing a practical approach to automated data formalization and indicating the importance of accessible, high-quality structured data for reliable downstream artificial intelligence applications.
Source: Nature Communications
Institution: Huazhong University of Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $6,990
Awards: G234177690, G6404540560, G776347602
OpenAlex ID: W7236584860
