Multi-Agent LLM Framework with Physics-Based Verification Boosts Structural Design Compliance to 98.6%

A new framework for code-compliant structural design using multi-agent large language models (LLMs) has been introduced in a paper on arXiv (2608.07978). The framework addresses the limitation of existing LLM systems that generate designs in one shot without verification, making them unsuitable for safety-critical tasks. Instead of relying on LLM self-correction, the proposed system injects feedback from an external physics-based verifier into a closed repair loop. The framework couples a three-layer finite-element verification system with a dual-node loop: Node 1 converts code violations into hard repair constraints, while Node 2 transforms a four-dimensional quality score into safety-first soft constraints. A retrieval-augmented code base ensures every violation is traceable to a specific clause. In tests across five structure types and 44 cases, code compliance improved from 56.8% to 98.6%, and the composite score rose from 63.8 to 71.4 (p<0.000001), while using about 5.8% less material. Removing either node degrades performance, and compliance did not change detectably across the two backbone LLMs tested, indicating that the improvement is attributed to the verification loop rather than the model choice. This work is relevant to the intersection of AI, engineering, and architecture, offering a method to enhance safety and efficiency in structural design.