Where to Edit, What to Create: CoMesh for AI-Augmented 3D Modeling by Beginners

Professional 3D modeling software is inaccessible to beginners: steep learning curves and dense interfaces demand expertise most users lack. Recent LLM agents enable natural-language control of 3D modeling, yet their prompt-only interaction leaves three usability gaps. First, users cannot inspect what the agent will do before its code modifies the scene. Second, changes clear in direction but vague in amount ("make this rounder") are hard to convey. Third, precise locations, such as a face, edge, or corner, are hard to name. To address these gaps, we present CoMesh, a mixed-initiative paradigm powered by multimodal LLMs that repositions the user from passive prompt-writer to active collaborator: agent intent is made visible and revisable before code runs, and vague or spatial changes are resolved through direct manipulation. In a within-subjects study (N=12 beginners), CoMesh completed tasks 48% faster and with 55% fewer corrections than a prompt-only baseline, and raised satisfaction to 6.53/7.