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

arXiv cs.GR details AgenticCADedit, a stateful, tool-mediated approach to multimodal 3D CAD editing that treats each change as a committed step in a persistent session. Instead of rebuilding a full CAD program from scratch after every attempt, the system can inspect faces and edges, highlight selected regions for verification, and revert only the operation that went wrong.

That matters for any workflow where artists or technical artists are iterating on existing geometry from speech, sketches, or other multimodal prompts. The core shift is from stateless regeneration to incremental editing, which preserves partial progress and gives the model feedback on what it just produced. In practical terms, that should reduce the waste that comes from repeatedly redoing nearly-correct edits.

The paper reports improvements across all metrics for three evaluated LLMs: qwen3.6-27b, gemma4-31b, and gpt-5.6-luna. The biggest jump was on qwen3.6-27b, where validity rose from 51.0% to 94.8% and acceptance from 1.6% to 12.0%.

There is also a cost angle: with gpt-5.6-luna, the authors report 66.7% fewer output tokens than neuralCAD-Edit, while 94.8% of input tokens came from the prompt cache. For teams thinking about AI-assisted DCC or CAD-style tooling, the interesting takeaway is that statefulness and tool feedback can matter as much as raw model size.