A Self-Verification Framework Toward Reliable Text-to-BIM Generation

Abstract: Recent advances in generative AI have enabled Text-toBIM systems that generate building models from naturallanguage prompts; however, these systems typically lack automated verification, requiring manual inspection and iterative error correction. To improve the reliability of LLM-generated BIM outputs, we propose a Text-toBIM workflow with integrated self-verification. For each prompt, the system derives two complementary validations: an Information Delivery Specification (IDS) for rule-based checking, and LLM-driven code-based verification for requirements beyond IDS. The resulting feedback is used to refine the IFC model in a closed loop. The framework is implemented and evaluated through a case study, demonstrating improvements in model quality