Research map / Large language models / Reasoning and chain of thought

From Intent to Execution: Multimodal Chain-of-Thought Reinforcement Learning for Precise CAD Code Generation

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 1 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 CAD-RL, a CoT cold-start plus RL post-training framework (executability, IoU, and GPT-4o rewards, with Trust Region Stretch, Precision Token Loss, Overlong Filtering) for generating CadQuery code from text/reference images, and releases the ExeCAD dataset of 16,540 examples.

Problem

Translating natural-language or structured design intent into executable, numerically precise CAD (CadQuery) code is hard because it requires logical reasoning, syntactic validity, and accurate dimensions, and existing VLMs and datasets fall short.

Approach

Combines CoT-annotated supervised cold start with RL post-training using a multiplicative executability gate plus IoU and LLM-judge rewards, and adds three modifications to PPO/GRPO: relaxed clip bounds (Trust Region Stretch), upweighted numeric tokens (Precision Token Loss), and dropping truncated samples from the reward (Overlong Filtering).

Key results

• CAD-RL reaches 78.7% IoU and 98.83% executability under structured design language input on ExeCAD, versus 70.4% IoU and 95.44% executability for CAD-Coder.
• Under natural language input, CAD-RL reaches 72.3% IoU and 99.63% executability, versus 66.2% and 94.72% for CAD-Coder and 47.3% IoU / 73.16% executability for GPT-4o.
• Ablation: RL post-training is the largest gain, raising natural-language IoU from 63.9% to 72.3% over the CoT+reference-image SFT variant.

Surprising findings

• Cold-start SFT alone already reaches about 90% executability but only about 50% IoU, showing executability and geometric accuracy diverge.

Limitations

• Remaining dimensional imprecision: IoU stays at 78.7% under structured input.
• Reliance on GPT-4o both for dataset alignment and as the external evaluation reward judge.
• Precision Token Loss weighting is described heuristically with no specified values for numeric-token weights.

Experimental setup

Datasets: ExeCAD GenCAD-Code Text2CAD

Benchmarks: ExeCAD

Metrics: IoU Mean Chamfer Distance Median Chamfer Distance executability

Baselines: CAD-Coder GPT-4o Claude4 Gemini2.5 Pro InternVL3 Qwen2.5-VL InternVL2.5 Phi-3.5-Vision LLaVA-1.5

Builds on: PPO GRPO DAPO-style overlong filtering CAD-Coder GenCAD-Code Text2CAD

Closest papers by meaning

• CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward (NeurIPS 2025 · similarity 0.92)
• cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.91)
• ReCAD: Reinforcement Learning Enhanced Parametric CAD Model Generation with Vision-Language Models (AAAI 2026 · similarity 0.90)
• Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.87)
• CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.86)
• CReFT-CAD: Boosting Orthographic Projection Reasoning for CAD via Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.86)