acceptodds
Under review as a conference paper at ICLR 2027

CAD-WAM: Anticipating Geometry for Text-to-CAD Generation

13 views

Abstract

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CAD generation, text-to-CAD, large language models, world models, 3D generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_pihwOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CAD-WAM, a text-to-CAD generation framework that combines autoregressive CAD operation generation with geometric state prediction. The method encodes intermediate CAD geometries into latent representations using a pretrained VAE and jointly predicts the next CAD operation and its resulting geometric state through a shared LLM backbone. During inference, generated operations are executed through a CAD kernel, and the resulting geometry is used to guide subsequent generation. The authors construct stepwise training trajectories from DeepCAD and evaluate the method against task-specific CAD models and general-purpose LLMs.

Critical strengths
  1. Well-motivated and technically sound approach. The proposed framework is coherent and addresses the limitations of autoregressive CAD generation by incorporating geometric feedback and next-state supervision.
  2. Strong experimental performance. CAD-WAM achieves convincing improvements over competitive baselines across multiple geometric metrics, particularly on longer construction sequences.
  3. Comprehensive experimental evaluation. The paper includes comparisons against recent CAD generation methods and general-purpose LLMs, along with reasonable ablation studies demonstrating the effectiveness of the proposed components.
Critical weaknesses
  1. Novelty. Although the framework is technically sound, several of its core components appear closely related to existing research. cad-specific tokenization and command-sequence modeling have been extensively explored in previous works, including Text2CAD, FlexCAD, and CAD-Tokenizer. Similarly, using VAE-based representations to encode geometry is a relatively established technique. Overall, I find the methodological contributions somewhat incremental, and the paper would benefit from a clearer justification of its novelty.
  2. Limited experimental scope. The evaluation is conducted exclusively on DeepCAD, which primarily contains relatively simple sketch-and-extrude operations. Given the increasingly saturated performance on this benchmark, additional experiments on more challenging datasets would strengthen the paper's claims regarding generalizability.
  3. Lacking qualitative examples of text-to-cad generation.

Overall, I think it is a purely borderline work. The content is quite solid, but the contribution is incremental and I am not sure if I genuinely learned something other than "this work is the best on benchmark xxx". It would make perfect sense and be a clear accept if it was evaluated by a second-tier conference, but I am just not sure that the requirements of insights and novelty are the same for ICLR. I am giving a 2 for initial score and set my confidence low, and would be very happy to hear from the authors.

Questions for authors
  1. Can the authors provide more qualitative samples? Specifically, the input prompts should be included with the demonstrations.
  2. How would the authors place the work among all the previous works? Where is the novelty coming from?
Reviewer_4MKnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CAD-WAM generates CAD models from text through executable modeling actions, using a Qwen2.5-1.5B backbone conditioned on the description and a SetVAE encoding of the current geometry. An auxiliary head predicts the next geometric state during training, while inference uses CAD-kernel execution to update the geometry after each action. Experiments on DeepCAD compare the method with task-specific models and general-purpose LLMs, with ablations examining state representations and next-state supervision.

Critical strengths
  1. The framework connects action generation with the geometry produced at each modeling step. Using executed solids as feedback and next-state prediction as auxiliary supervision is an innovative touch that avoids relying on predicted geometry during inference.

  2. CAD-WAM achieves the lowest mean Chamfer distance in every construction-length group, while maintaining a low invalidity rate as the number of actions increases. These results support its reliability on longer construction sequences, where errors can accumulate across modeling steps.

  3. This paper is well written and easy to follow.

Critical weaknesses
  1. The descriptions list the construction steps with exact dimensions, and the prompt also gives the number of completed actions (Figure 9), so much of the next action can be read directly from the text. Neither an ablation without the geometric state nor the accuracy of the predicted state is reported, leaving open how much the model actually relies on geometry.

  2. The task-specific baselines use the textual annotations they were trained on, whereas CAD-WAM uses Recap-OmniCAD descriptions. Differences in the level of detail between these inputs may contribute to the gap with these baselines, making it hard to isolate the method’s advantage.

  3. Next-state supervision reduces mean Chamfer distance from 1.316 to 1.142, while the median changes only slightly. A paired analysis of per-model errors and confidence intervals would help determine whether the improvement is broadly distributed or mainly driven by a few large-error cases.

Questions for authors
  1. How does performance change when the current geometric state is removed while keeping the textual description and progress information unchanged?

  2. How much of the performance gap could be explained by differences in the information provided in the textual descriptions?

Reviewer_ghxBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method to perform text-to-CAD generation from textual descriptions. CAD-WAM combines a JSON-based DSL that bundles a sketch, extrusion and boolean operation into one action with a latent representation of the current geometry. A fine-tuned LLM generates the next action, while an auxiliary head predicts the next geometric state. At inference, each action is executed by a CAD kernel and the resulting solid is re-encoded for the next step. The evaluation compares the framework with existing text-to-CAD methods on DeepCAD and includes ablations of the state representation and prediction head.

The topic is relevant, and I find learning from geometric outcomes an appealing idea. However, the experiments do not sufficiently establish the magnitude of the contribution. The DSL appears to be a relatively straightforward repackaging of existing modeling operations, while the evidence for the added value of future-state prediction is limited. My main concern is whether the experiments support a substantial contribution beyond the combination of these components.

Critical strengths
  • Future-state prediction appears beneficial, reducing mean Chamfer distance by 13.2% in the ablation.
  • The complete framework achieves strong geometric accuracy on DeepCAD.
Critical weaknesses

Clarity of Exposition

The architecture and execution loop are generally understandable. However, the distinction between predicting a desired next state and simulating the consequences of an action needs more discussion. In Section 3.4, the prediction head reads the hidden states at the geometric input positions, before the action tokens are generated. Its input contains the description and current state, but not a candidate action. Therefore, with the description and current geometry fixed, it cannot predict different outcomes for different candidate actions. I read this as auxiliary supervision of the next state rather than an action-conditioned simulator. This does not make the objective invalid, but it weakens the connection to counterfactual reasoning and planning that motivates world models.

I also find the novelty of the DSL limited. Bundling a sketch profile, extrusion and boolean operation into a JSON object seems like a fairly simple representation choice. Pointer-CAD already groups sketch-and-extrude combinations into complete steps and additionally supports chamfer and fillet through geometric entity pointers, which the proposed DSL lacks. CADmium already uses JSON-based CAD sequences. The comparison in Section 3.2 between a complete action and a single primitive token or code statement does not establish that existing representations cannot provide intermediate solids at completed-operation boundaries. The paper should explain what the proposed DSL adds beyond this grouping and why a new language is preferable to FeatureScript or CadQuery.

Validation

The evaluation covers several geometric and validity metrics and shows encouraging performance for the complete framework. However, it does not convincingly isolate the contribution of the proposed DSL or establish a strong benefit specifically from predicting future geometric states.

For the DSL, I would like a controlled comparison using the same backbone, CAD examples, textual descriptions, geometric feedback and training objectives, with the output represented in the proposed JSON DSL, DeepCAD/Text2CAD-style commands grouped into completed operations, FeatureScript and CadQuery. The complete-system comparisons in Table 1 change several of these factors simultaneously. They therefore do not show whether the new DSL accounts for the improvement. A matched action-only model without geometric input or state supervision would also help separate the contribution of the action representation from the geometry components.

The evaluation should also include more complex CAD histories. DeepCAD is restricted to sketch-and-extrude programs, and the converted dataset contains one to ten modeling actions, averaging about two actions per model from the counts in Appendix D. This setting offers limited evidence for the practical advantages of a new DSL or for anticipating geometry over longer sequences. Omni-CAD/Recap-OmniCAD provides a concrete next experiment with more complex sketch-and-extrude histories that could be converted into the current action language.

For a richer operation set, CADFS provides a native text-conditioned benchmark of real-world Onshape designs represented in FeatureScript, including operations such as revolve, loft and fillet. Comparing with CADFS in that setting would be more convincing than evaluating its FeatureScript model only on DeepCAD. CadQuery-based methods such as cadrille have also been evaluated beyond DeepCAD, including on real-world CC3D data, which motivates a broader comparison with this code representation.

For future-state prediction, Table 3 reduces mean Chamfer distance from 1.316 to 1.142, a 13.2% improvement, which is encouraging. However, median Chamfer distance changes only from 0.167 to 0.166, IoU from 0.960 to 0.964, and invalidity from 0.779% to 0.666%. I find these gains modest overall for the paper's central additional learning objective. The strong results of the complete pipeline against external baselines cannot by themselves establish the importance of the prediction head, since the variant without this head already performs well.

I would like to see direct evaluation of the predicted next states. The VAE reconstructions in Appendix B do not evaluate the LLM's forecasts. Comparing next-state prediction with current-state reconstruction or final-state supervision would also clarify whether anticipating the next state is important, or whether another geometric auxiliary objective gives similar improvements. Finally, the head ablation should be broken down by construction length and difficult cases. A more challenging dataset with longer histories would make this test much more informative. Future prediction might support better geometric representations, action checking or planning, but the current experiments demonstrate only a limited auxiliary-training benefit, and the head's lack of action conditioning further limits the simulator interpretation.

Explanation of Rating

I lean toward a weak reject. The problem is relevant, the framework is reasonable, and the reported improvements are encouraging. However, I do not find that the experiments sufficiently validate the amount of contribution claimed. The DSL has limited novelty and flexibility, and lacks controlled comparisons with existing representations on richer CAD data. Future-state prediction is potentially valuable, but its specific benefit is not convincingly established by the modest aggregate ablation gains, the restricted dataset and the absence of direct forecast evaluation or alternative auxiliary-objective controls. Stronger experiments isolating these two contributions would substantially strengthen the paper.

Questions for authors
  • What advantage does the proposed DSL provide over prior command representations, FeatureScript or CadQuery when the backbone, training examples, geometric feedback and learning objectives are matched? Can this advantage be demonstrated on more complex CAD histories beyond DeepCAD and potentially in-the-wild data?

  • How accurately does the head predict next states, particularly for operations that substantially change the geometry? Does next-state supervision outperform other geometric auxiliary objectives, such as current-state reconstruction or final-state prediction?

  • Are the ablation gains consistent across training runs and construction lengths? A breakdown on longer histories and an analysis of the failure cases improved by state prediction would help clarify its practical benefit.

Reviewer_ekXUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAD-WAM, which encodes sampled point clouds into latent states and jointly trains an LLM for next-action generation and next-state prediction. During inference, each generated action is executed by a CAD kernel, and the resulting geometry is re-encoded to guide the next step. The authors report higher geometric accuracy than the evaluated baselines on DeepCAD, with ablation results indicating benefits from auxiliary next-state supervision.

Critical strengths
  • The paper is well-written and easy to follow.
  • The work has potential practical value, as generating editable CAD models from natural-language descriptions could reduce manual modeling effort.
Critical weaknesses
  • The evaluation is limited to DeepCAD, with training and testing conducted in an in-distribution setting. It remains unclear whether the reported improvements generalize to other datasets or out-of-distribution CAD models.
  • The paper does not report results across multiple independent training runs or provide standard deviations or confidence intervals. This makes it difficult to assess the stability of the reported results and improvements.
Questions for authors
  • Regarding the textual-state baseline in Section 5.3: Is the textual description sufficiently detailed to preserve the spatial information needed for CAD modeling? How does it capture face boundaries and connectivity, and could information loss in this representation contribute to the performance gap in Table 2?
  • What motivates the use of point clouds over other spatial representations, such as meshes or voxels? Could you provide comparisons to better justify this choice?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.