Year: 2026
Type: conference-paper
Abstract: The growing reliance on software in safety-critical systems demands strong guarantees of functional correctness. Formal methods address this need by providing proofs as evidence of correctness. Among these, Correctness-by-Construction (CbC) ensures functional correctness through the incremental application of refinement rules while providing early feedback on the program correctness. However, its practical adoption is limited by the required manual effort. Among others, writing concrete program code is left to the developer. In recent years, large language models (LLMs) have demonstrated strong capabilities in code generation from both natural language and formal inputs. Therefore, we explore the use of LLMs to automatically generate code in the context of CbC to make the approach more applicable. We evaluate LLM-based program synthesis using four LLMs and a prompt tailored to CbC on a dataset of 57 specifications drawn from existing CbC case studies, each consisting of a precondition and a postcondition describing the desired behavior. We assess the generated code with respect to correctness and reproducibility across multiple runs. In addition, we compare the results to classically synthesized code, as it can currently be generated in CbC. Our evaluation shows that all considered LLMs outperform the current synthesis approach. Among them, Claude achieves the best performance, reaching 80.70% correct results.
Institution: Karlsruhe Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $5,392
OpenAlex ID: W7204971070