Research map / Multi-agent systems and games / LLM agents

Clarify Before You Draw: Proactive Agents for Robust Text-to-CAD Generation

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ProCAD, a two-agent system where a proactive clarifying agent asks targeted questions about ambiguous CAD prompts before a CadQuery coding agent generates the program, both fine-tuned from Qwen2.5-7B-Instruct.

Problem

Text-to-CAD prompts are often under-specified or internally inconsistent (missing dimensions, conflicting constraints), and existing one-shot fine-tuned models hallucinate dimensions instead of resolving ambiguity.

Approach

Decompose text-to-CadQuery into a proactive clarifying agent that audits the prompt and asks minimal clarification questions (two-round policy, trained via agentic SFT on synthetic ambiguous-prompt trajectories) followed by a coding agent that turns the corrected specification into CadQuery code; the coding agent is trained on a curated 10K text-to-CadQuery dataset with descriptions generated by G

Key results

• ProCAD reduces mean Chamfer distance by 79.9% versus Claude Sonnet 4.5 and lowers invalidity ratio from 4.8% to 0.9% on the ambiguous test set (two-agent setting with ProCAD-coder).
• ProCAD-coder reaches 0.9% invalidity ratio on 1K unambiguous prompts versus 12.9% for Claude 4.5 Sonnet.
• In human-centered evaluation, ProCAD-clarifier scores 0.760 efficiency and 0.787 resolution versus 0.598 and 0.700 for Claude Sonnet 4.5.

Surprising findings

• Zero-shot Qwen2.5-7B-Instruct has ~85% invalidity mostly from CadQuery syntax errors, yet 1.6K SFT samples bring it to 0.9% invalidity.
• Claude Sonnet 4.5 often fails to detect genuine under-specification and proceeds without clarifying.

Limitations

• User responses are simulated by GPT-5-mini, so real-user behavior is only partially tested (small 100-example human study).
• Assumes the simulated user answers any clear question correctly, leading to a fixed two-round policy.
• Dataset descriptions depend on a closed-source VLM (GPT-5-mini) and ambiguity types are limited to under-specified and conflicting dimensions.

Experimental setup

Datasets: DeepCAD CAD-recode CadQuery dataset

Benchmarks: ProCAD ambiguous-prompt test set (2,469 samples) ProCAD unambiguous test set (1K samples) Text2CAD subset

Metrics: Chamfer distance invalidity ratio efficiency F1 resolution score

Baselines: Claude Sonnet 4.5 GPT-4o-mini Qwen2.5-7B-Instruct Cadrille Text-CAD JSON-Distill

Builds on: Text2CAD CAD-recode self-refine

Closest papers by meaning

• CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward (NeurIPS 2025 · similarity 0.89)
• Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts (NeurIPS 2024 · similarity 0.89)
• CAD-Editor: A Locate-then-Infill Framework with Automated Training Data Synthesis for Text-Based CAD Editing (ICML 2025 · similarity 0.88)
• Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.87)
• CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.86)
• Text-to-CAD Evaluation with CADTests (NeurIPS 2026 · similarity 0.86)