@inproceedings {pub6392,
	title = {Human-in-the-Loop: Quantitative Evaluation of 3D Models Generation by Large Language Models},
	author = {Ahmed Sadik AND Mariusz Bujny},
	year = {2026},
	month = {March},
	abstract = {Recent advancements in Large Language Models (LLMs) have demonstrated their potential in generating 3D shapes, yet the evaluation of these outputs remains a critical challenge. In this study, we propose a quantitative evaluation framework that integrates human-in-the-loop feedback to iteratively refine 3D shape generation. Our approach employs a set of complexity and similarity metrics to assess the generated models against ground truth (GT) counterparts. We introduce Surface Complexity, Feature Complexity, and Topological Complexity to measure the structural intricacy of the generated 3D shapes. Additionally, we assess alignment and similarity through Principal Component Analysis (PCA) Alignment Score, Hausdorff Distance, Iterative Closest Point (ICP) Alignment Score, Dimensional Similarity, Volumetric Similarity, and Surface Similarity. By providing designers with quantitative validation, our framework enables iterative refinements in the generative process, ultimately improving the quality and fidelity of LLM-generated 3D models. This research bridges AI-generated creativity with human design intuition, fostering a collaborative paradigm for intelligent shape synthesis.},
	publisher = {Springer, LNCS, LNAI, LNBI},
	booktitle = {18th International Conference on Agents and Artificial Intelligence }
}