acceptodds
Under review as a conference paper at ICLR 2027

SAR4CAD: Structure-Aware Reinforcement Learning for Parametric CAD Reconstruction from Multi-View Drawings

Abstract

Computer-Aided Design (CAD) models can be reconstructed from engineering drawings; however, existing methods are limited by the scarcity of well-annotated drawing datasets and an excessive reliance on a single sequence representation. Specifically, most sequence generators optimize only one serialization scheme, although neither sketch ordering nor the decomposition of certain features is necessarily unique. SAR4CAD is a two-stage framework that trains a multimodal large language model (MLLM) using multi-view renderings, scalable vector graphics (SVG) projection tokens, and concise textual descriptions. The SVG tokens encode visible and hidden edges, primitive types, and quantized coordinates extracted from the CAD model, thereby providing reliable geometric information for reconstruction. In the first stage, supervised fine-tuning teaches the model the required output syntax and establishes a stable generation policy. In the second stage, structure-aware group relative policy optimization (GRPO) evaluates sampled outputs after they have been parsed and executed. The associated reference-guided structural reward integrates CAD-kernel validity, structural-count similarity, sketch overlap computed via Hungarian matching, ordered extrusion-parameter similarity, and 3D Chamfer similarity. Experiments on a constructed parametric CAD dataset demonstrate that SAR4CAD consistently outperforms the evaluated baselines across all evaluation metrics. Code is available at bluehttps://anonymous.4open.science/r/SAR4CAD-F64F.

Primary area:
reinforcement learning
Keywords:
CAD Reconstruction, Reinforcement Learning, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gfjoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Finetuned some model on image --> CAD task, where image is a multi-view image (top / left / bottom / etc), and the output is a full parametric 3D CAD.

Small innovation on forcing the generated CAD program to be parsable with a syntax check (strcture aware ..)

Optimization is rather hard as CAD structure is quite discrete, so you cannot back-prop gradients, but you need some kind of RL to handle these discrete choices, which the authors do.

Critical strengths

It works well on the things they tested.

Critical weaknesses

They don't really test it on more complicated geometries, it seems the complexity of the generated CAD are <30 faces, or simpler. Only 3 extrusions or so.

Lack of a gallery. Papers like these at least need a 1 page "gallery" showing concrete datapoints of what the model can do. For example, if your paper is on image generation, have a gallery of 100 images generated by your neural network. If this work is img --> CAD, I'm expecting at least 20+ instances of image --> CAD examples, displayed proudly in the main text of the paper.

The fact that the authors did not display what they could generate proudly leads me to conclude they're worried that their actual results are not impressive.

Which also leads me to suspect if I just prompt GPT-astra and it would just 1 shot their task.

Questions for authors

Do baseline comparison against GPT astra. For a maximum budget of 100 dollars, at least.

Reviewer_WsUjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SAR4CAD, a framework that reconstructs sketch–extrude CAD sequences from multi-view drawings using a multimodal LLM. In addition to four raster views, the model receives a short text description and SVG edge tokens that encode primitive type, visibility, and quantized coordinates. Training proceeds in two stages: supervised fine-tuning to learn the output grammar, then GRPO with an execution-gated composite reward that combines validity, structural counts, Hungarian-matched sketch overlap, ordered extrusion similarity, and 3D Chamfer similarity. Experiments on a DeepCAD-derived dataset compare the method with zero-shot Qwen2.5-VL and Drawing2CAD, and include ablations of reward terms, input modalities, and training stages.

Critical strengths
  1. Each view is accompanied by SVG tokens that encode primitive types, visible and hidden edges, and quantized coordinates on the same grid as the output sequence. Adding these tokens to images and text substantially reduces typical-case reconstruction error, showing that explicit geometric conditioning is a simple and effective complement to raster inputs.

  2. Invalid programs are filtered by the CAD kernel, and sketches are matched regardless of order, so partially correct outputs still receive graded credit. Both progressive and leave-one-out ablations examine each reward term, and the code is released.

  3. SVG tokens provide primitive types, edge visibility, and quantized coordinates alongside the raster views. Their addition substantially reduces median reconstruction error in the supervised setting, supporting the value of explicit geometric conditioning.

Critical weaknesses
  1. The added value of the structural rewards is not clearly established. Reinforcement learning with geometric rewards has already been applied to CAD reconstruction, e.g., in cadrille (Kolodiazhnyi et al., 2026), but this work is neither discussed nor compared. The ablations also lack a variant that uses only validity and 3D shape similarity under the same training setup, so it remains unclear how much the structural terms contribute beyond simpler geometric feedback.

  2. At both training and test time, the SVG tokens with visibility labels are extracted directly from the ground-truth CAD solid, and the textual descriptions are Text2CAD annotations of the same model (Sec. 3.3). These inputs provide clean geometric and semantic information that would not be available in practice, and no real engineering drawings are tested. The results therefore do not establish how well the method works with the “readily available drawings” discussed in the introduction.

  3. Figure 1(b) motivates the method using independent features whose order can change without affecting the final solid. However, extrusion blocks are matched by position, and the extrusion and joint terms together carry 55% of the reward weight (Sec. 3.5). Although positional matching is reasonable for order-dependent operations, it may still penalize equivalent reorderings. The paper does not quantify this penalty or show whether the method learns valid alternatives to the reference sequence.

Questions for authors
  1. How does the full reward compare with a baseline that uses only validity and 3D shape similarity (r_3D) under the same training setup? This comparison would clarify the added value of the structural terms.

  2. For equivalent sequences that reorder independent features, as in Figure 1(b), how much does the total reward decrease relative to the reference sequence?

  3. How does SAR4CAD perform without the textual description, i.e., with inputs closer to those of Drawing2CAD? This would clarify how much of the advantage over Drawing2CAD comes from the method rather than from the additional input.

1 comment
Reviewer_WsUjOpenReview ↗

Response to authors' feedback

Thank you for the response, which has addressed some of my concerns. However, I still believe that the current experiments do not provide strong enough support for the authors’ claims, so I find it difficult to raise my score to 3. I would consider a score of 2.5, but since this option is not available, I will keep my original score.

I also carefully read the other two reviewers’ comments and the authors’ follow-up responses. Unfortunately, the authors have not provided the additional experimental comparisons requested by these reviewers, such as comparisons with ReCAD and GPT-6 Astra. I therefore believe that the paper still needs further work, particularly in its experimental design and evaluation.

Reviewer_odHLOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SAR4CAD for parametric CAD reconstruction using multi-view drawings, SVG tokens, and structure-aware reinforcement learning. Experiments on a DeepCAD-derived dataset show improvements over Drawing2CAD and the supervised-only variant.

Critical strengths
  1. The paper recognizes the limitations of supervising CAD generation with a single reference sequence and explores structure-aware rewards to address them, with ablations supporting improvements in validity and mean geometric accuracy.
Critical weaknesses
  1. Comparing only with Drawing2CAD and an untuned Qwen model is insufficient. Without comparisons to existing CAD reinforcement learning methods, such as cadrille and ReCAD, the paper’s additional contribution remains unclear.
Questions for authors

None

Comments on the paper

1 comment

Thanks again for authors' further response. The new discussion looks good to me so I decided to raise the score accordingly. If the paper is accepted, these results and analyses should be included in the final version.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.