acceptodds
Withdrawn from ICLR 2027

Geometry Before Symbols: Latent State-Transition Modeling for Autoregressive CAD Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent image-to-CAD methods increasingly formulate CAD reconstruction as autoregressive generation of executable feature sequences. However, the next feature is predicted in symbolic sequence space, while its semantics are ultimately defined by the geometric state transition it induces. We introduce ForeCAD, a framework for latent geometric state-transition modeling in autoregressive CAD generation. Before generating each feature, ForeCAD forms internal representations of the current construction state and the anticipated feature-induced transition, so that the symbolic feature is generated as the realization of an internal geometric decision. We ground these representations with execution-derived supervision of intermediate states, material additions and removals, and state–transition consistency. We further introduce Representation-Guided Execution Alignment (RGEA), which aligns generated feature consequences with the geometric intent encoded by the learned representations. To support feature-aligned supervision and evaluation, we construct FeatureState-CAD, spanning Procedural Construction and DeepCAD-Complex. ForeCAD consistently outperforms sequence-based and geometry-aware baselines; on DeepCAD-Complex, it improves IoU from 36.86% to 39.53%, execution rate from 88.25% to 90.00%, and mean CD from 0.090 to 0.079. Controlled ablations and representation interventions further validate the learned latent state–transition representations and the benefit of aligning their geometric intent with executable feature generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer-Aided Design, Large Vision-Language Models, Reinforcement Learning, Geometry-Grounded Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fWZmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ForeCAD, an autoregressive framework for CAD program generation that explicitly models latent geometric states and transitions before generating CAD features. It introduces geometric supervision for state and transition representations and a Representation-Guided Execution Alignment (RGEA) strategy optimization stage. Experiments on Procedural Construction and DeepCAD-Complex show improvements over several baselines in geometric reconstruction and execution success.

Critical strengths

1.The paper addresses the important challenge of generating executable CAD programs that produce geometrically accurate shapes.

2.The explicit geometric supervision and execution-based alignment provide a structured approach to incorporating geometric information into CAD generation.

3.The experiments evaluate both geometric reconstruction quality and program execution success.

Critical weaknesses
  1. Necessity of the proposed formulation. Both ForeCAD and an execution-feedback-based agent aim to improve CAD generation using geometric information. However, the latter can directly execute CAD operations and use the resulting geometry to guide subsequent decisions. The paper does not sufficiently justify why explicitly predicting latent geometric transitions is preferable to this alternative.

  2. Generalization to complex objects. The current benchmarks do not sufficiently establish whether ForeCAD generalizes to substantially more complex geometric structures beyond the patterns represented in its training data. Evaluations on more diverse and structurally challenging objects would strengthen the paper.

  3. Contribution of the transition representation. It remains unclear how accurately predicting latent geometric transitions translates into better discrete CAD feature generation. The authors should provide more direct evidence linking transition prediction quality to final reconstruction performance.

Questions for authors

See weakness

Reviewer_ESo8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ForeCAD, a framework for image-to-CAD generation that addresses the mismatch between symbolic feature prediction and its geometric consequences. ForeCAD introduces latent representations of the current geometric state and anticipated state transition to guide autoregressive CAD generation. These representations are supervised using execution-derived geometric information and further optimized through Representation-Guided Execution Alignment (RGEA), a reinforcement learning procedure. The authors also construct FeatureState-CAD, a dataset with feature-level geometric annotations. Experiments demonstrate improvements over several baselines.

Critical strengths
  1. The paper is generally well structured, with a coherent formulation and detailed methodological explanations.
  2. The authors provide ablations of the main components and representation interventions to investigate the effectiveness of the proposed approach.
Critical weaknesses
  1. The motivation and methodology overlap with existing research on CAD representations and geometry-aware generation. For instance, CAD-Tokenizer (ICLR 2026) has explored CAD-specific representations beyond conventional language tokens, and cadrille (ICLR 2026) has investigated execution-based geometric feedback and reinforcement learning for CAD reconstruction. Although ForeCAD introduces a different formulation, its distinctions and contributions relative to these works are insufficiently established.
  2. Several relevant methods, particularly cadrille, are missing from the experimental comparisons without adequate justification. The baseline selection is not fully plausible for me. I do not understand how good the baselines selected by the authors are, and therefore feel hard to compare this work with others I know.
  3. Even with the selected baselines, the improvement over the strongest baseline is relatively modest. Notably, ForeCAD without RGEA underperforms the strongest geometry-aware baselines, suggesting that the improvements rely heavily on reinforcement learning rather than the proposed latent representations. The paper does not sufficiently establish the benefits of its geometric modeling compared with simpler execution-guided optimization under matched training budgets.
  4. Direct visual comparisons with competing methods are missing.
Questions for authors
  1. related works: how would the authors compare their work with previous works as cad tokenizer and cadrille in terms of motivation and contribution?
  2. baselines: how do works like cadrille, cad coder work on the benchmarks proposed by this work?

Honestly, I feel the authors need to review the related works in this area and find good, strong baselines to compare with. At least they need a sound justification for not doing so. The current form of the paper is solid, but to me it does not bring a lot of new knowledge. I gave a 2 but my actual position is around 1.5, and I feel hard to change my position unless there are new, strong experimental results and explanations.

Reviewer_SvEaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ForeCAD, a latent geometric state-transition framework for autoregressive CAD generation motivated by the mismatch between transition in semantic space and geometric space. It is equipped with a Representation-Guided Execution Alignment (RGEA) technique, which aligns generated feature consequences with the geometric intent encoded by the learned representations. It also constructs a dataset, FeatureState-CAD, where ForeCAD outperforms sequence-based and geometry-aware baselines.

Critical strengths

The paper is generally well structured. The identification of the semantic and geometric mismatch is somewhat valuable.

Critical weaknesses

The performance improvements are marginal (e.g., IoU increasing from 36.86% to 39.53% and execution rate from 88.25% to 90.00%). This raises concerns over whether moving from the symbolic space to the latent space yields genuine benefits, or if these gains stem merely from minor implementation details—especially given that all baselines were re-trained by the authors on their own FeatureState-CAD dataset, as noted in the caption of Table 1.

It remains unclear how the central issue highlighted in Figure 2's caption—that "equal symbolic edits can induce unequal geometric consequences"—is resolved or mitigated by ForeCAD, either quantitatively or qualitatively. The reported evaluation metrics do not directly reflect or measure this issue.

The presentation lacks clarity, particularly in Section 3.4, which lacks sufficient detail and concrete explanations.

Questions for authors

What do L_CAD, L_S, and L_P specifically represent in the objective in equation 9 exactly? What is a field value y in equation 7? What is the core motivation behind Representation-Guided Execution Alignment (RGEA)? Could the authors elaborate on its design and intuition in more detail? What is the primary motivation for introducing the FeatureState-CAD dataset over existing benchmarks?

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

Related papers

Discussion (0)

Sign in to comment.