From Intent to Execution: Multimodal Chain-of-Thought Reinforcement Learning for Precise CAD Code Generation
TLDR
Proposes CAD-RL, a multimodal Chain-of-Thought reinforcement learning framework for precise CAD code generation from natural language, with a new real-world dataset ExeCAD.
Reasoning
The paper introduces a novel combination of CoT reasoning and RL with task-specific rewards for CAD code generation, supported by a new dataset of real-world CAD examples. Strengths include clear methodology and optimization strategies; weaknesses are that the abstract is cut off, limiting full assessment of results and limitations.
Read-first score
Read-first score 63.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 84.
Field roles
Rank sensitivity
Stability: volatile; rank range: 8.
Keyword Scores
Deep Analysis
Innovations
- Multimodal Chain-of-Thought guided reinforcement learning post-training framework for CAD code generation
- Three task-specific rewards: executability reward, geometric accuracy reward, and external evaluation reward
- Three optimization strategies for stable policy learning: Trust Region Stretch, Precision Token Loss, and Overlong Filtering
- ExeCAD dataset: 16,540 real-world CAD examples with paired natural language, structured descriptions, executable CADQuery scripts, and rendered 3D models
Methodology
CAD-RL combines Chain-of-Thought cold start with goal-driven reinforcement learning post-training using three rewards (executability, geometric accuracy, external evaluation). To handle sparse and high-variance rewards, it introduces Trust Region Stretch, Precision Token Loss, and Overlong Filtering. The ExeCAD dataset supports training and benchmarking.
Key Results
CAD-RL achieves significant improvements in reasoning quality, output precision, and code executability over existing vision-language models.