PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models
TLDR
PR-CAD unifies text-to-CAD generation and editing via progressive refinement with LLMs, achieving state-of-the-art controllability and faithfulness.
Reasoning
The paper's strength lies in unifying generation and editing tasks with a reinforcement learning-enhanced framework and a curated interaction dataset. Weaknesses include limited detail on dataset size and specific benchmark metrics in the abstract.
Read-first score
Read-first score 50.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 59.
Field roles
Rank sensitivity
Stability: volatile; rank range: 24.
Keyword Scores
Deep Analysis
Innovations
- Unified progressive refinement framework that combines generation and editing for text-to-CAD
- High-fidelity interaction dataset covering full CAD lifecycle with multiple representations and qualitative/quantitative descriptions, defining edit operations and human-like interaction data
- Reinforcement learning-enhanced reasoning framework integrating intent understanding, parameter estimation, and precise edit localization into a single agent
- CAD representation tailored for LLMs
Methodology
PR-CAD uses a progressive refinement framework with a reinforcement learning-enhanced agent that integrates intent understanding, parameter estimation, and edit localization. It is trained on a curated dataset spanning the CAD lifecycle with multiple representations and both qualitative and quantitative descriptions, using a CAD representation designed for LLMs.
Key Results
Experiments show mutual reinforcement between generation and editing tasks and across modalities; PR-CAD achieves state-of-the-art controllability and faithfulness on public benchmarks in both generation and refinement, and significantly improves CAD modeling efficiency.