AgenticCADedit: Streamlined 3D CAD Editing for Designers

Paper: AgenticCADedit: A Stateful, Tool-Mediated Agentic Approach to Multimodal 3D CAD Editing

Authors: Saptarshi Neil Sinha, Mika Silvan Goschke, Paul Julius K"uhn, Arjan Kuijper, Michael Weinmann

Executive Summary

The research presents a new approach to computer-aided design (CAD) called AgenticCADedit, which addresses limitations in existing systems for editing CAD models. Traditional methods often require designers to regenerate entire programs from scratch for each editing request, leading to inefficiencies where partially correct work is lost and making it difficult to track changes or revert mistakes.

AgenticCADedit improves this process by allowing for incremental editing. Instead of generating a complete design in one go, it breaks down editing into a series of small, verifiable actions that build upon each other. This means that each step commits to the current state of the design, enabling designers to inspect the outcome of their modifications immediately. If an error occurs, they can revert that specific action without losing all previous work.

The research assessed the performance of AgenticCADedit against traditional methods using three different large language models (LLMs): two open-weight models (qwen3.6-27b and gemma4-31b) and one proprietary model (gpt-5.6-luna). The results demonstrated significant improvements across all models. For instance, the weakest baseline model, qwen3.6-27b, saw its validity rate jump from 51.0% to 94.8%, and its acceptance rate increased from 1.6% to 12.0%. These metrics indicate that AgenticCADedit not only enhances the accuracy of design edits but also improves acceptance by users.

Additionally, a cost analysis of token usage revealed that the proprietary model (gpt-5.6-luna) produced about 66.7% fewer output tokens compared to traditional editing methods. This efficiency could lead to reduced computational costs and faster processing times, which are crucial factors for industrial applications where time and resources are limited.

The implications of this research are significant for industries that rely on CAD for manufacturing and design. By facilitating a more efficient and flexible editing process, firms could potentially reduce the time spent on model adjustments and improve overall design quality. This technology may be particularly beneficial for sectors such as automotive, aerospace, and consumer electronics, where precision and rapid iteration are essential.

While the results come from simulations and benchmark tests rather than real-world applications, the advancements suggest a promising direction for integrating AI into CAD workflows. AgenticCADedit could represent a valuable tool for designers looking to enhance their productivity and accuracy in complex design tasks.

Academic Abstract

Computer-aided design is central to industrial manufacturing, and much of a designer's daily work consists of editing existing models from multimodal requests involving speech, sketches, and model interaction. Existing neural CAD approaches focus predominantly on unconditional or text-conditioned generation. The neuralCAD-Edit approach formalizes expert multimodal editing requests, but its iterative baseline refines a complete CAD program across attempts, executing each attempt from the original model in a stateless CAD environment. Every attempt must therefore reconstruct the entire edit from scratch, so partially correct progress is discarded rather than accumulated, and the model can neither inspect the geometry it has just produced nor selectively revert a single faulty operation. We present AgenticCADedit, which turns editing into a sequence of small, verifiable actions on a persistent CAD state instead of a single regenerated program. Rather than emitting one complete program, it applies incremental code steps that each commit to the session, inspects the resulting faces and edges, renders highlighted selections to verify that the intended region was addressed, and reverts individual operations when it was not. Subsequent actions therefore build on the geometry produced by earlier ones. Our approach improves on all metrics for all three evaluated LLMs (open-weight: qwen3.6-27b, gemma4-31b; proprietary: gpt-5.6-luna), with the largest gains for the weakest baseline model, qwen3.6-27b, whose validity rises from 51.0% to 94.8% and acceptance from 1.6% to 12.0%. A token-cost analysis with gpt-5.6-luna further shows $66.7$% fewer output tokens than neuralCAD-Edit, while $94.8$% of input tokens are served from the prompt cache.