aDSL: Agentic 3D Creation via Joint Agent-Program Design

TL;DR - aDSL jointly designs a 3D-focused domain-specific language and a role-specialized multi-agent workflow to make LLM-generated 3D programs more reliable, controllable, and editable. It improves text-to-shape and image-to-shape performance by replacing brittle coordinate choices with semantic structure and spatial relations.

• The DSL provides composable relational operators for expressing geometry and constraints rather than relying on absolute coordinates.
• A training-free Plan-Execute-Critic loop decomposes requests, generates code, and repairs errors or constraint violations using execution feedback.
• Experiments report better robustness and intent fidelity than prior LLM-based baselines while retaining explicit structure and interpretability.
• The approach also supports articulated-object creation and structured scene composition.