2026 Volume 7 Issue 3 Pages 1085-1094
Automated structural inspection of bridges using LiDAR point clouds is expected to contribute to inspection cost reduction and safety improvement, but existing multimodal large language models (LLMs) tend to systematically misanswer structural facts such as the number of piers and spans. In this study, we propose BridgeLLM, which connects a point cloud encoder BridgeSPT integrating adaptive voxelization, PCA geometric descriptors, additive geometric conditioning, serialized window attention, and member recognition summary tokens to Qwen3-VL-8B-Instruct, and introduce structural grounding that injects structural facts from deterministic geometric analysis into prompts. Evaluation results on 124 out of 616 QA cases from 10 real bridges showed that the structural fact accuracy (SFA) improved from 0.815 to 0.887, and pier counting improved from 0.25 to approximately 1.0.