Research map / Vision-language and multimodal / Multimodal LLMs

ReCAD: Reinforcement Learning Enhanced Parametric CAD Model Generation with Vision-Language Models

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 3 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.

ReCAD fine-tunes Qwen2.5-VL-7B on parameterized CAD code and then applies GRPO with code-guided off-policy rollouts and hierarchical primitive curriculum to generate parametric CAD models from text or images.

Problem

Generating precise, editable parametric CAD models from text or image inputs, where prior methods rely on SFT knowledge injection, generalize poorly out-of-distribution, and produce imprecise parameters.

Approach

Rewrite hardcoded CAD scripts into parameterized code with a VLM (filtered by DINOv2 similarity), use that code to generate descriptions for SFT, then run RL (GRPO) where parameterized code guides rollouts for hard questions, with a curriculum over primitive hierarchy and a reward combining IoU and image-feature similarity.

Key results

• Image-to-CAD on DeepCAD: mean CD reduced from 73.47 (CAD-Coder) to 29.61, and on Fusion 360 OOD from 272.06 to 80.23.
• Text-to-CAD on DeepCAD: mean CD 72.47 vs 98.12 for CAD-LLaMA, with P-F1 61.48 vs 60.02.
• Ablation on DeepCAD text-to-CAD: removing guidance raises mean CD from 72.47 to 87.34; removing HPL raises it to 90.83.

Surprising findings

• Zero-shot editing and debugging emerge without task-specific training, according to the authors.
• CAD-LLaMA has lower invalidity ratio than ReCAD-VL on DeepCAD text-to-CAD (0.39 vs 0.81).

Limitations

• Failure cases include mismatches with input descriptions and spatial parameterization errors.
• Relies on GPT-4o for code rewriting and description generation, and on in-house SFT data not released.

Experimental setup

Datasets: DeepCAD Fusion 360 Gallery UltraChat OpenCodeReasoning Text2CAD descriptions CAD-LLaMA descriptions

Benchmarks: DeepCAD Fusion 360 Gallery

Metrics: P-F1 Chamfer Distance (median, mean) Invalidity Ratio IOU_best

Baselines: GPT-4o DeepSeek-V3 Qwen3-235B-A22B CAD-Translator Text2CAD CAD-LLaMA CAD-Coder Gemini-2.5-Pro Qwen-2.5-72B

Builds on: GRPO CAD-Coder CAD-LLaMA DeepCAD Text2CAD

Closest papers by meaning

• Revisiting CAD Model Generation by Learning Raster Sketch (AAAI 2025 · similarity 0.91)
• From Intent to Execution: Multimodal Chain-of-Thought Reinforcement Learning for Precise CAD Code Generation (AAAI 2026 · similarity 0.90)
• cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.90)
• CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward (NeurIPS 2025 · similarity 0.88)
• From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.87)
• Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models (ICML 2025 · similarity 0.87)