Rethinking Scientific Modeling: Toward Physically Consistent and Simulation-Executable Programmatic Generation
Researchers have developed a new framework to improve the generation of scientific modeling code using large language models (LLMs). This approach integrates domain knowledge, aligns models with constraints, and uses verification for evaluation. A new dataset called CivilInstruct and a two-stage fine-tuning strategy were introduced to ensure the generated code is physically consistent and executable for simulations, significantly reducing errors compared to existing methods. AI
IMPACT Enhances the reliability and applicability of LLMs in scientific and engineering domains by ensuring code consistency and executability.