LLM-based Visual Code Completion for Aerospace Geometric Design

Researchers have built a visual code-completion copilot for aerospace design using a large language model, but slow inference times restrict its practical use to complex, time-consuming tasks, according to a new paper [1]. The system, described in a preprint submitted to arXiv on June 15, 2026, applies a visual programming variant of the ReAct methodology and is powered by GPT 5.4 [1]. It is paired with Wingbuilder, a new plugin library for the Grasshopper visual programming environment that provides custom components for aerospace-specific geometry abstraction [1]. The work also introduces the Aerospace Visual Programming Dataset, or AVPD, containing 18 expert-designed tasks at varying difficulty levels alongside ground-truth solutions [1]. Large language models are a type of machine learning model trained on vast text corpora for tasks such as language generation [10]. Their application in safety-critical industries has moved cautiously. The aerospace sector, which prioritizes safety and explainability over rapid adoption, currently has no publicly announced LLM-based geometric design copilot systems in commercial use by original equipment manufacturers [1]. To gauge real-world viability, the researchers conducted a user trial with two experienced aerospace engineers from a large aircraft manufacturing company [1]. The copilot successfully generated suggestions that participants found helpful, and the paper states that “participants reported they liked the tool and would be willing to use it in the future” [1]. However, the slow inference times of the ReAct loop meant the tool was only considered worthwhile for more complex tasks where waiting for a quality suggestion was justified [1]. The submission file weighs 4,674 KB and was uploaded by Stuart Middleton [1]. The paper appears on arXiv under the Computation and Language category and is indexed on Hugging Face’s daily papers feed, where community members can link models, datasets, and interactive demos to the preprint [6][7][8]. Broader AI adoption in engineering continues to expand. Artificial intelligence encompasses capabilities such as learning, reasoning, and decision-making, with the subfield of machine learning already deployed in language translation, image recognition, and e-commerce [3]. Within aerospace specifically, quantitative methods like the Datar–Mathews method for real options valuation, created in 2000 with contributions from a Boeing Technical Fellow, illustrate the industry’s long-standing integration of computational decision-support tools [4]. The new copilot represents a step toward bringing generative AI into that lineage, though the inference-speed bottleneck remains a barrier to routine use [1].