IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing
TLDR
IterCAD is a multimodal agent for closed-loop, interactive CAD generation and editing, outperforming existing methods in code executability and geometric precision.
Reasoning
The paper presents a novel iterative framework with a data synthesis pipeline and geometry-aware reinforcement learning, addressing the gap between one-shot generation and iterative real-world practices. Strengths include a comprehensive evaluation suite and strong empirical results, but the abstract lacks explicit details on limitations or comparisons to specific baselines.
Read-first score
Read-first score 58.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 80.
Field roles
Rank sensitivity
Stability: volatile; rank range: 5.
Keyword Scores
Deep Analysis
Innovations
- Unified multimodal agent framework (IterCAD) for closed-loop, interactive CAD generation and editing via multi-turn interaction with an executable CAD sandbox
- Data synthesis pipeline that generates standard-compliant multi-view engineering drawings, complex code-editing tasks, and high-fidelity interaction trajectories incorporating advanced industrial manufacturing features
- Progressive supervised fine-tuning followed by geometry-aware reinforcement learning with viable-prefix masking to enhance code executability and geometric fidelity
- IterCAD-Bench evaluation suite and the Chamfer Distance Tolerance-Recall (CD-TR) curve with AUC-TR metric, providing a survivor-bias-free standard that unifies code validity and geometric precision
Methodology
IterCAD frames CAD generation and editing as multi-turn interactions between a multimodal agent and an executable CAD sandbox across three tasks: Drawing-to-Code, Text-to-Code, and Interactive Editing. A data synthesis pipeline creates training data with multi-view engineering drawings and interaction trajectories, and the agent is optimized via progressive supervised fine-tuning followed by geometry-aware reinforcement learning with viable-prefix masking. Evaluation uses the new IterCAD-Bench and the CD-TR curve with AUC-TR metric.
Key Results
IterCAD achieves highly competitive performance, significantly outperforming existing approaches in both code executability and geometric precision, and demonstrates superior closed-loop iterative refinement capabilities.