Driving · 2026-10-30
CAD-WAM: Anticipating Geometry for Text-to-CAD Generation
Zhengyang Li, Weijie Zheng, Yuxin Cui, Sijie Wang, Wenjie Feng, Qiyu Kang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
CAD-WAM generates text-to-CAD models by predicting both the next executable CAD-DSL action and the next geometric state. It encodes intermediate solids as latent states from surface point clouds, and uses an auxiliary state-prediction head for geometric supervision during training. At inference, each predicted action is executed by a CAD kernel and the resulting solid is re-encoded for the next decision, improving geometric accuracy and validity on DeepCAD.
Why it matters
The authors address multi-step text-to-CAD reliability by explicitly supervising what intermediate geometry each action should produce, rather than only predicting commands. Their closed-loop kernel execution with re-encoded geometric feedback connects action generation to geometric outcomes, and their ablations isolate the impact of next-state supervision on geometry quality.
Method
- Define CAD modeling operations as executable actions in a CAD DSL; use CAD kernel execution to deterministically transition from an intermediate solid to the next solid.
- Encode intermediate solids by sampling surface point clouds and using a pretrained SetVAE encoder to produce compact latent geometric states for both conditioning and state-prediction supervision.
- Use an LLM backbone with two heads: an Action Head to generate the next DSL action and a State Prediction Head to regress the next latent state; during inference, execute predicted actions and re-encode the resulting solid.
Limitation
The evaluation is limited to sketch-and-extrude programs in DeepCAD. Extending CAD-WAM to richer operations and longer construction sequences would help assess its applicability to more demanding modeling tasks.
Abstract (from arXiv)
Generating editable computer-aided design (CAD) models from text requires selecting operations that produce the intended changes to the geometry constructed so far. An incorrect operation can compromise subsequent steps, making reliable multi-step generation challenging. Existing methods based on large language models (LLMs) primarily learn to predict commands, without explicit supervision of the geometric states those commands should produce. Inspired by world-action models, we introduce CAD-WAM, which couples geometry-conditioned action generation with next-state prediction. We define complete modeling operations as actions in a CAD domain-specific language (DSL) and encode surface point clouds of intermediate solids into compact latent states using a pretrained variational autoencoder. An LLM backbone generates the next action from the target description and current geometric state, while an auxiliary head on the same backbone predicts the next state to provide geometric supervision during training. We construct step-wise training trajectories from parametric CAD models in the DeepCAD dataset, pairing each trajectory with an expert-level textual description. At inference, each generated action is executed by a CAD kernel, and the resulting solid is re-encoded to inform the next decision. On the DeepCAD test set, CAD-WAM outperforms the task-specific baselines in geometric accuracy and output validity, with larger gains on models requiring more operations, and produces more accurate geometry than general-purpose LLMs prompted with the same DSL. Ablations further show that next-state supervision reduces mean Chamfer distance by 13.2\% relative to an otherwise identical model trained only to predict actions. Code is available at https://github.com/Lzy-1037/CAD-WAM.
Related papers
- E0.5: From Reconstructive to Future-Predictive 3D Foundations for Robotic World-Action Modeling
- GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation
- XGenAct: Geometry-Enhanced World Action Models through Cross-Task Generation
- Relational Geometry Forcing for World Action Models
- Action Jets: Learning the Local Shape of the Action-conditioned Data Manifold
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC
CAD-WAM: Anticipating Geometry for Text-to-CAD Generation
Preprint posted: 2026-10-29 · Original record ↗
- arxiv
- 2610.39306
Explicit source versions (1)
- v1 · 2026-10-29
- Preprint available
Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.39306v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Generating editable computer-aided design (CAD) models from text requires selecting operations that produce the intended changes to the geometry constructed so far. An incorrect operation can compromise subsequent steps, making reliable multi-step generation challenging. Existing methods based on large language models (LLMs) primarily learn to predict commands, without explicit supervision of the geometric states those commands should produce. Inspired by world-action models, we introduce CAD-WAM, which couples geometry-conditioned action generation with next-state prediction. We define complete modeling operations as actions in a CAD domain-specific language (DSL) and encode surface point clouds of intermediate solids into compact latent states using a pretrained variational autoencoder. An LLM backbone generates the next action from the target description and current geometric state, while an auxiliary head on the same backbone predicts the next state to provide geometric supervision during training. We construct step-wise training trajectories from parametric CAD models in the DeepCAD dataset, pairing each trajectory with an expert-level textual description. At inference, each generated action is executed by a CAD kernel, and the resulting solid is re-encoded to inform the next decision. On the DeepCAD test set, CAD-WAM outperforms the task-specific baselines in geometric accuracy and output validity, with larger gains on models requiring more operations, and produces more accurate geometry than general-purpose LLMs prompted with the same DSL. Ablations further show that next-state supervision reduces mean Chamfer distance by 13.2\% relative to an otherwise identical model trained only to predict actions. Code is available at https://github.com/Lzy-1037/CAD-WAM.
