acceptodds
Under review as a conference paper at ICLR 2027

HistCAD: Constraint-Aware Parametric CAD Histories for Evaluating Editability

Abstract

Sketch constraints specify geometric conditions for constructing and modifying parametric CAD models. We study whether predicted constraints allow a given history to reproduce the required initial model and support prescribed dimensional edits. We introduce HistCAD, an executable representation and dataset whose Academic and Industrial collections contain 180,495 parametric construction histories with entity-referenced sketch constraints and retained feature operations. Predictors receive these histories with the geometry and feature definitions retained and explicit sketch constraints removed. They generate constraints for every sketch without seeing the edit request. The benchmark compares models built with alternative constraint sets for the same history under the same dimensional edit. An edit succeeds when the model reproduces the required initial geometry, reaches the target value, preserves specified relations and unedited dimensions in the target sketch, and rebuilds through the complete history. A predictor trained on both collections and supplied with descriptions of the input histories achieves overall edit success of 52.4% on Academic and 29.0% on Industrial. Models retaining only endpoint-connectivity constraints in the target sketch and the history's constraints elsewhere can reach the target and rebuild while failing preservation. For all-sketch predictions, we retain the target-sketch prediction and restore the history's constraints in other sketches. More models then reproduce the required initial geometry, and some of these newly matched models complete the edit. HistCAD connects constraint learning to the construction and revision of parametric CAD models.

Primary area:
datasets and benchmarks
Keywords:
parametric CAD, sketch constraints, CAD editability, construction histories, constraint prediction, executable evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_76gmOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

HistCAD presents a dataset containing 180,495 parametric construction histories along with code to import and manipulate them in Fusion 360. The paper presents metrics which measure how well the parametric models retain design intent when edited. Withholding the sketch constraints is shown to decrease the metrics. A Qwen3-8B model is also fine-tuned to generate the models, with parametric history, based on text descriptions.

Critical strengths

It’s great to see a paper which combines both constrained parametric sketches and the 3D features built from them. Most previous works have created “dumb tool bodies” or predicted sketch geometry and constraints, but never studied how well these constraints capture design intent when edited. HistCAD is the first study I have seen which combines parametric sketches, 3D modelling and studies how design intent is preserved as the sketch parameters are varied.

The dataset of 180,495 parametric construction histories is massive benefit to the community.

Removing the DeepCAD duplicated boundaries (Appendix A.2) is super useful. I always wanted to write code to do that!

Critical weaknesses

The paper is badly written and very hard to understand. It looks like the authors might have run out of space and moved too much information into the appendices. The main paper doesn’t link to the appendices and the whole thing is very hard to follow.

The paper says “We introduce an executable representation” but this is not really described. It sounds like you use Fusion 360 to do the actual parametric modelling. Given a sketch on a planar face of a previous extrusion, would the sketch plane move if the planar faces moved (i.e. do we have inter-feature constraints)? If so, how are these inter-feature references stored? Perhaps too much critical detail was moved into appendices and the main paper no-longer makes sense.

Figure 2 isn’t helping me understand the metrics. The text around this is quite hard to parse as well. It would be good to see examples of cases where the pre-edit match, T, P and V are failing. Understanding the metrics becomes very hard. Looks like there is something in the appendix, but I’m not sure I understand.

Text-conditioned inputs are documented in Appendix A5. This needs to be referenced from main paper. It would be useful to see some examples in the Appendix.

It’s not clear if the results in Table 2 are related to the fine tuned model or experiments with the dataset.

Questions for authors

Description–history correspondence

I don’t follow why the “shuffled history” gives such high metric numbers. 20% is OES, even though the model was prompted with a different text description. Is there no geometry matching check? Why isn’t Pres low?

Target-sketch controls, Table 3

I’m struggling to follow this section

  • Are you reporting results for the Qwen3-8B output or the training dataset?
  • The text which tries to explain the results doesn’t add any understanding beyond what we have in the table. Are you saying something like “When we don’t use the constraints then 64% of the models are deformed in a way which doesn’t retain the design intent”?
  • I would have thought that OES would fall to zero without constraints. Perhaps there are lots of pairs of concentric circles which the coincident constraints cover It sounds like section 7 is trying to give some explanation of this. After reading it carefully I still don’t have a clue what it’s trying to say

Section 6.2 Table 4 I don’t understand at all. We need images showing an example of what is being modified

With the target-sketch prediction fixed, restoring the history’s constraints in other sketches enables additional models to reproduce the required initial geometry I can’t parse this sentence. Are you saying you fix one sketch and predict the rest of the history? Is the prediction with the network or using the parametric recipe?

Reviewer_3GMEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HistCAD, a representation and dataset of parametric CAD histories containing sketch geometry, entity-referenced constraints, and feature operations. The task is to predict constraints for all sketches while retaining the geometry and construction history and removing explicit sketch constraints. Predictions are evaluated through initial-model reconstruction and prescribed dimensional edits. Text descriptions serve as optional auxiliary inputs, while editing is performed by a fixed evaluation procedure.

Successful editing requires matching the initial geometry, attaining the requested dimension, preserving specified relations and unedited dimensions, and successfully rebuilding the complete construction history. The authors establish Qwen3-8B baselines and examine the effects of training data, textual descriptions, and constraints outside the target sketch. The best setting achieves overall edit success rates of 52.4% on Academic and 29.0% on Industrial. The results show that high constraint F1, target attainment, or successful rebuilding alone does not guarantee successful editing.

Critical strengths
  1. Evaluating constraints through actual construction and editing behavior.
    The paper goes beyond matching predicted constraints against reference entries by requiring correct initial reconstruction, target attainment, preservation of specified relations, and successful rebuilding. The Connectivity control and high-F1 failure cases demonstrate that constraint matching or execution success alone can overlook consequential errors. This provides an evaluation approach more closely connected to practical CAD use and constitutes the paper’s strongest contribution.

  2. Controlled experiments identify the influence of constraints throughout the construction history.
    By keeping the target-sketch prediction fixed and restoring the original constraints in other sketches, the authors observe improvements in initial matching and overall edit success. The additional successful cases originate from tasks that previously failed initial matching. This experiment demonstrates why constraints should be evaluated within complete construction histories rather than solely within the sketch being edited. Together with the stage-specific metrics, this design helps explain failure causes.

Critical weaknesses

Insufficient specification and behavioral validation of the conversion from source CAD models to HistCAD.

The paper represents parametric models from different sources using a finite, unified constraint vocabulary. However, it does not sufficiently explain how driving versus driven dimensions, dependencies expressed through parameter formulas, unsupported source constraints, and constraint transfer after primitive splitting or merging are handled. For example, replacing a parameter dependency with its current numerical value may preserve initial geometry while changing subsequent editing behavior.

The experiments primarily construct the Full reference using the converted history’s own constraints. They therefore evaluate predictions relative to this reference but do not independently establish that conversion preserves the editing semantics of the source model. This does not directly invalidate evaluation within the unified representation, but it limits confidence in its fidelity to the original parametric histories. For a paper whose central contributions include a representation and dataset, this is a substantive validation gap.

Questions for authors
  1. How are driven or reference dimensions, parameter-formula dependencies, and unsupported source constraints handled during conversion? Could these transformations change the editing behavior permitted by the original model?

  2. Have paired editing tests been performed on source models and their HistCAD counterparts? If so, please report geometric and rebuilding consistency under identical dimensional edits. If not, please clarify which properties the current validation establishes and which it does not.

Reviewer_ZUAWOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The paper presents the HistCAD history dataset, containing 180,495 parametric modeling histories and 3,476,914 sketch constraints, and establishes an evaluation pipeline for constraint prediction and parametric editing. The model predicts missing constraints while preserving the initial sketch geometry and feature operations; the evaluation checks for consistency in initial geometry, achievement of target dimensions, retention of specified relationships, and successful reconstruction of the entire modeling history. Jointly trained models using the same history description achieve overall editing success rates of 52.4% on Academic histories and 29.0% on Industrial histories.

Critical strengths

Critical Strengths

(1) A practical evaluation objective and experimental setup; (2) A composite dataset of open-source CAD histories; (3) Additional industry histories;

Critical weaknesses

(1) The overall pipeline remains largely similar to those of prior methods such as Pointer-CAD, Seek-CAD, EPCCAD and other existing pipelines.

Dacheng Qi, Chenyu Wang, Jingwei Xu, Tianzhe Chu, Zibo Zhao, Wen Liu, Wenrui Ding, Yi Ma, Shenghua Gao. Pointer-CAD: Unifying B-Rep and Command Sequences via Pointer-based Edges & Faces Selection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 17377-17387.

Xueyang Li, Jiahao Li, Yu Song, Yunzhong Lou, Xiangdong Zhou. Seek-CAD: A Self-refined Generative Modeling for 3D Parametric CAD Using Local Inference via DeepSeek. International Conference on Learning Representations 2026 (ICLR 2026)

Yin Z, Lin W, Kong X. A multimodal benchmark for editable constraint preserving history based CAD modeling. Discover Artificial Intelligence. 2026 Sep 2;6(1):1022. (EPICCAD)

(2) In certain aspects, HistCAD even falls short of the above methods.

For example of the selection, one of the key techniques for Evaluating Editability.

  1. Chamfer { "near_points": [ [0.0, 10.0, 10.0] ], "plane": [1.0, 0.0, 0.0], "dist": 1.5, "angle": 30.0, "operation": "Chamfer" }

near_points: Boundary selection markers. Proximity 3D sampling points used to resolve the targeted edges for the chamfer operation.

HistCAD uses the point coordinate [0.0, 10.0, 10.0] to select the edge to be chamfered.

When the scope of parametric edit is small, the target edge may still be selected, and the subsequent chamfer operation is therefore carried out successfully.

However, when the scope of parametric edit is large, the point coordinate [0.0, 10.0, 10.0] no longer selects the edited edge, and the subsequent chamfer operation cannot be performed.

This kind of selection with the point coordinate [0.0, 10.0, 10.0] is too simple to evaluate editability when being compared with Pointer-CAD and Seek-CAD.

Since both the parametric edit and the selection operation are too simple, the benchmark for evaluating editability is of little significance.

(3) For example of the constraint, one of the key techniques for Evaluating Editability.

The author still avoids any comparison with EPCCAD.

Questions for authors

(1) The overall pipeline remains largely similar to those of prior methods such as Pointer-CAD, Seek-CAD, EPCCAD and other existing pipelines.

Dacheng Qi, Chenyu Wang, Jingwei Xu, Tianzhe Chu, Zibo Zhao, Wen Liu, Wenrui Ding, Yi Ma, Shenghua Gao. Pointer-CAD: Unifying B-Rep and Command Sequences via Pointer-based Edges & Faces Selection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 17377-17387.

Xueyang Li, Jiahao Li, Yu Song, Yunzhong Lou, Xiangdong Zhou. Seek-CAD: A Self-refined Generative Modeling for 3D Parametric CAD Using Local Inference via DeepSeek. International Conference on Learning Representations 2026 (ICLR 2026)

Yin Z, Lin W, Kong X. A multimodal benchmark for editable constraint preserving history based CAD modeling. Discover Artificial Intelligence. 2026 Sep 2;6(1):1022. (EPICCAD)

(2) In certain aspects, HistCAD even falls short of the above methods.

For example of the selection, one of the key techniques for Evaluating Editability.

  1. Chamfer { "near_points": [ [0.0, 10.0, 10.0] ], "plane": [1.0, 0.0, 0.0], "dist": 1.5, "angle": 30.0, "operation": "Chamfer" }

near_points: Boundary selection markers. Proximity 3D sampling points used to resolve the targeted edges for the chamfer operation.

HistCAD uses the point coordinate [0.0, 10.0, 10.0] to select the edge to be chamfered.

When the scope of parametric edit is small, the target edge may still be selected, and the subsequent chamfer operation is therefore carried out successfully.

However, when the scope of parametric edit is large, the point coordinate [0.0, 10.0, 10.0] no longer selects the edited edge, and the subsequent chamfer operation cannot be performed.

This kind of selection with the point coordinate [0.0, 10.0, 10.0] is too simple to evaluate editability when being compared with Pointer-CAD and Seek-CAD.

Since both the parametric edit and the selection operation are too simple, the benchmark for evaluating editability is of little significance.

(3) For example of the constraint, one of the key techniques for Evaluating Editability.

The author still avoids any comparison with EPCCAD.

2 comments

Q1. Research contribution and the “overall pipeline” comparison

HistCAD provides a representation, dataset, and executable benchmark to study whether sketch constraints support dimensional editing while preserving specified relations and unedited dimensions. Two models can have identical initial geometry and still behave differently under the same edit. For example, two lines may initially form a right angle in both models, yet remain perpendicular in only one after a length change. Reconstructing the initial shape does not establish whether the constraints will preserve the required relationships during editing.

HistCAD represents sketch geometry, explicit constraints, and feature operations within an executable construction history. Each constraint identifies the sketch entities or points it acts on, such as the two lines required to remain perpendicular. The dataset provides 180,495 such histories from Academic and Industrial collections. Its stored constraints supply training targets for constraint prediction, while the associated geometry and feature operations allow the predictions to be tested through model construction and editing. Because the constraint sets are stored separately, we can replace them while keeping the stored geometry definitions, feature operations, and feature references fixed. The representation and dataset thus let us test different constraint sets on the same construction history.

In the prediction experiments, we remove all explicit sketch constraints from an existing history and ask Qwen to predict a constraint set for every sketch. Qwen receives the stored geometry and feature operations, but not the later edit request. Its output specifies constraints on the existing sketch entities or points. Fusion 360 then solves the sketches with those constraints and executes the stored feature operations. The learned constraints are therefore tested on a subsequent edit that the predictor has not seen.

Different predicted constraints can change the solved geometry even when the stored geometry definitions are unchanged. The benchmark first checks whether the reconstructed model matches the initial geometry built with the stored constraints. For a matching model, the evaluator applies a prescribed dimension change and checks whether the target is reached, the specified relations and unedited dimensions are preserved, and every feature operation executes successfully. This separates reproducing the initial model from preserving its required behavior after an edit.

Pointer-CAD generates CAD commands using B-Rep-based entity selection, while Seek-CAD studies feedback-driven CAD generation. HistCAD compares how models built from the same construction history with different explicit sketch-constraint sets respond to the same prescribed edit. Similarities in using CAD histories or executing feature operations do not establish that these works evaluate this constraint-preserving behavior.

Q2. Point-based selection and the significance of the benchmark

The conclusion that the benchmark is “of little significance” does not follow from the chamfer example. A chamfer’s edge reference determines where the feature is applied. A sketch constraint determines which relationships the sketch solver must maintain as dimensions change. HistCAD evaluates both complete-history execution and preservation of the required sketch relationships; successful execution alone is not treated as successful editing.

The Full–Connectivity experiment separates these outcomes without using predicted constraints. Full retains all stored sketch constraints. Connectivity keeps only the Coincident constraints connecting profile endpoints in the target sketch; all other sketches retain Full’s constraints. These constraints keep the connected endpoints together, but do not require two lines to remain perpendicular. Both conditions start from matched initial geometry. The evaluator applies the same dimension change to both, adding the dimension constraint if it is absent.

The edit specifies the dimension to change; the preservation checks specify what should stay unchanged. For both variants, the evaluator checks the same selected relations and unedited dimensions from the stored target sketch. Removing a constraint changes what the solver enforces, but does not remove the corresponding preservation requirement. The evaluator measures the resulting geometry directly, so a relation can pass even when no corresponding explicit constraint is present.

Figure 1 makes this distinction concrete. The two lines are initially perpendicular in both variants. After the same length edit, both models reach the target and rebuild successfully, but only Full keeps the lines perpendicular. Connectivity fails the preservation check because the edited geometry has lost the required right angle, despite attaining the requested length and executing all feature operations. The failure concerns the model’s response to the edit, not an operation’s inability to find an edge.

The quantitative results show that successful rebuilding does not imply passing the preservation checks. On Industrial, 412 of 500 Connectivity edits reach the requested dimension and successfully execute the complete history. Of these, only 183 pass every preservation check, leaving 229 edits that rebuild but fail preservation checks (Section 6.1; Table 3).

In the chamfer example raised in the review, the original query-point coordinates are reused after the edit. If the edge reference cannot be resolved, the feature cannot execute and the evaluator records a rebuilding failure (Section 3.1; Appendix A.2). That execution failure cannot explain these 229 outcomes, because their complete histories have already executed successfully. Even a single dimensional edit can therefore expose a failure that target attainment and successful rebuilding alone would miss.

Q3. The requested comparison

The cited paper was published on September 2, 2026. It falls within the contemporaneous-work period defined by the ICLR 2027 Reviewer Guidelines for peer-reviewed papers published on or after July 17, 2026. The guidelines state that authors are not required to compare with such work and that “the lack of such comparisons cannot be a basis for rejection.”

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

As the discussion period draws to a close, we would like to summarize HistCAD's main contributions, the additional evidence provided during the rebuttal, and the revisions made to the manuscript.

HistCAD establishes a foundation for studying how learned geometric relationships govern editing behavior in 3D parametric CAD. Its three main contributions are: (1) an executable representation connecting entity-referenced sketch constraints with retained 3D feature histories; (2) an Academic–Industrial dataset of 180,495 construction histories containing 3,476,914 sketch constraints; and (3) an executable benchmark evaluating initial reconstruction, prescribed dimensional edits, geometric preservation, and complete-history rebuilding.

Our experiments demonstrate the significance of these contributions. On Industrial tasks, 55.6% of reachable Connectivity edits fail preservation despite successfully attaining their targets and rebuilding, showing that successful execution alone does not establish editability. We have also added independent native NX editing validation, achieving 93.2% post-edit geometric consistency for standard HistCAD and 99.0% for Parameterized HistCAD in a separate study.

We believe our responses and the revised manuscript have substantially addressed the central concerns about representation fidelity, evaluation methodology, and presentation clarity. We have clarified the learning task and experimental procedures, improved the manuscript's organization, and brought key CAD examples into the main text, with substantive revisions highlighted in blue.

We respectfully ask the reviewers and Area Chair to consider the revised manuscript and additional evidence when assessing HistCAD's contributions and significance.

Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.