Towards Proactive Text-to-CAD Generation

ACL ARR 2025 May Submission6604 Authors

Abstract: Computer-Aided Design (CAD) systems are indispensable in mechanical engineering and product development processes. Nowadays, text-to-CAD methods can significantly reduce the learning cost of complex CAD systems and have attracted increasing attention. However, such methods fail to achieve alignment among user expectations, textual descriptions and CAD models. To address this limitation, we propose a new paradigm, ''Proactive Text-to-CAD Generation'', which first employs large language models to proactively elicit and formulate text enriched with comprehensive CAD design details, then generates CAD models from these refined descriptions. To support this paradigm, we construct the first actively interactive text-to-CAD dataset, Proactive-Text2CAD, which contains 4,590 high-quality dialogues. Moreover, building upon this dataset, we propose a novel agentic framework for this task, named ''Proactive Agent'', which is driven by a hierarchical finite state machine accompanying with three carefully designed modules. Extensive evaluation and comprehensive analysis on the Proactive-Text2CAD dataset demonstrate the effectiveness of both our proposed paradigm and agentic framework, with our method achieving significant improvements in both textual detail refinement and final CAD model generation quality.

Paper Type: Long

Research Area: Special Theme (conference specific)

Research Area Keywords: Text-to-CAD, Agent, Proactive Conversational AI

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 6604