acceptodds
Withdrawn from ICLR 2027

Promptable Geometric Grounding in B-Rep Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We study the task of grounding in Boundary Representation (B-Rep) models, a foundational capability for AI-driven Computer-aided design (CAD) modeling. Such grounding provides a semantic interface for expressing user intent, yet remains largely unexplored. Unlike traditional 3D visual grounding that localizes objects in real-world point cloud scenes, the task requires both a fundamentally different scene representation and a finer localization granularity, where targets are not objects but topologically defined primitives (e.g., faces and edges) embedded in a structured geometric model. This paper introduces PGGM, a Promptable Geometric Grounding Model for B-Rep models. PGGM jointly supports natural language descriptions and partial selections as prompts. It decouples continuous geometry and discrete topology: pretrained VAEs encode geometric shape information, while a topology encoder captures structured primitive relations. These representations are then integrated through prompt-conditioned fusion and primitive-level decoding for fine-grained grounding. To further improve hard-negative discrimination, we introduce Discriminative Bootstrapping (DB), a training strategy that leverages the partial branch to turn noisy auxiliary signals into discriminative supervision. DB constructs complementary signals from external teachers, model self-bootstrapping, and proximity-based hard negatives, exposing the model to distractors that are semantically plausible, self-predicted, or close to the target in latent space. This enables more reliable grounding under fine-grained geometric ambiguity. Experiments show that PGGM significantly improves B-Rep grounding over strong baselines, achieving a +17.49% F1 gain and a +24.18% precision gain in the text-only setting.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
B-Rep, Geometric Grounding, CAD, Primitive-Level Selection, Geometry-Topology Learning, Discriminative Bootstrapping

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_78JZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies promptable geometric grounding in boundary representation (B-Rep) CAD models, aiming to identify the specific faces or edges that correspond to natural-language descriptions of geometric design intent. The authors propose Promptable Geometric Grounding in B-Rep Models (PGGM), a framework that integrates pretrained geometric representations, geometric attributes, B-Rep topology, and text-conditioned feature interaction to predict the relevant geometric elements. The paper further introduces Discriminative Bootstrapping (DB), which combines teacher-generated selection signals, model-prediction-based partial selections, and hard-negative mining to improve grounding performance. Experiments on the FutureCAD benchmark show that PGGM achieves an F1 score of 69.01, outperforming the strongest reported baseline at 51.75. The paper also reports ablation studies and a downstream text-to-CAD generation experiment, in which replacing the existing grounding module with PGGM improves several generation metrics. These results suggest that incorporating geometric and topological information together with language-conditioned selection can improve fine-grained grounding in CAD models.

Critical strengths
  1. The paper studies fine-grained grounding of natural-language modeling intent to faces and edges in B-Rep CAD models. This is a meaningful problem for interactive CAD modeling and text-to-CAD systems, where incorrect primitive selection can directly affect downstream operations.

  2. PGGM achieves an F1 score of 69.01 on text-conditioned grounding, outperforming the strongest reported baseline by 17.26 percentage points. The ablations suggest that Discriminative Bootstrapping contributes substantially to text-conditioned performance. The downstream text-to-CAD experiment also provides preliminary evidence that the proposed grounding module can improve practical generation metrics.

  3. The method combines geometric features, B-Rep topology, text conditioning, and partial-selection prompts. The paper is generally well organized and provides implementation details and analyses of failure cases.

Critical weaknesses
  1. The methodological novelty is not yet fully established. PGGM combines several established components, including pretrained geometric encoders, topology-aware message passing, prompt-conditioned attention, contrastive learning, and hard-negative training. Although the overall model performs well, the current experiments do not fully isolate whether the gains arise from the proposed architecture, the Discriminative Bootstrapping strategy, or their interaction. In particular, the DB ablation establishes the usefulness of the complete strategy but does not sufficiently compare it with strong alternative approaches such as standard hard-negative mining, self-training, or teacher-generated pseudo-labels under comparable training conditions. Stronger controlled comparisons would help clarify the method's contribution beyond the integration of existing techniques.

  2. The generalization evidence is limited relative to the claims about robust geometric grounding. The main evaluation is based on FutureCAD, supplemented by training data derived from ABC. The paper removes overlapping sample IDs, which is useful, but it remains unclear whether the evaluation adequately tests generalization to geometrically similar but distinct CAD models, unseen combinations of modeling operations, or substantially different prompt formulations. The reported failure cases also indicate that ambiguous descriptions and compositional spatial relations remain challenging. More systematic cross-distribution evaluation and analysis of actual downstream operation success would strengthen the evidence for the method's practical reliability.

Questions for authors
  1. Can the authors compare Discriminative Bootstrapping with strong alternative hard-negative mining, self-training, and teacher-supervision baselines under comparable data and compute budgets?
  2. How are train/validation/test splits constructed at the CAD-model level, and beyond removing overlapping sample IDs, are near-duplicate or geometrically similar models across splits explicitly checked?
  3. Can the authors evaluate generalization to unseen combinations of modeling operations or CAD models from a different source, and report performance by operation type and prompt complexity?
  4. How often do grounding errors lead to failed or invalid downstream CAD operations, and can the authors provide a more direct evaluation of operation execution success?
Reviewer_9zDjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PGGM (Promptable Geometric Grounding Model), a framework designed for primitive-level visual grounding in Boundary Representation (B-Rep) CAD models. Unlike traditional 3D visual grounding on point cloud scenes that localizes separable 3D objects with bounding boxes, B-Rep grounding requires identifying fine-grained parts—such as specific faces and edges—embedded within continuous geometric boundaries.

Critical strengths

Novel & Realistic Task Formulation: Fills a significant gap at the intersection of interactive CAD modeling and 3D visual grounding by enabling language and interaction-driven primitive selection directly on native B-Rep topology.

Principled Architectural Decoupling: Combining continuous surface/curve VAE representations with coedge graph message passing effectively models geometry-topology coupling without relying on lossy point cloud conversions.

Effective Hard-Negative Mitigation (Discriminative Bootstrapping): The DB training strategy systematically trains the model to distinguish true targets from semantically, self-predicted, or latent-space distractors, driving marked improvements in grounding precision

Critical weaknesses

Limited methodological novelty : The architecture is an assembly of established parts (B-repGen VAEs, B-RepNet coedge message passing, BERT, transformer cross-attention, prototype attention pooling). Discriminative Bootstrapping (DB) combines well-known ingredients: hard-negative mining (PCD), self-training on top-k predictions (SBD), teacher pseudo-labels (TID), and a consistency loss. DB is called "a general training strategy", but it is evaluated on one task and one dataset. The related work dismisses SAM-style i.i.d. prompt noise as improving only robustness, yet this is never tested, and DB is not compared against simpler alternatives (random noisy partials, standard hard-negative mining, plain self-training). Since the partial branch is unused at text-only test time, DB's text-only gain (+8.5 F1) may reflect extra training signal per sample rather than the specific DB design.

Lack of Ablation: The architectural components (VAE features, gate, topology branch, geometry bias) are not ablated at all, so the "decoupled geometry-topology" claim is also untested

Clean Partial Prompt Performance Trade-off: Presented ablation studies show that training with Discriminative Bootstrapping slightly reduces peak performance on clean partial-only inputs in exchange for significantly improved robustness under text-conditioned and noisy environments

Questions for authors

Can you ablate the VAE features, topology branch, and geometry bias to support the geometry-topology decoupling claim?

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

This paper extends text-based B-Rep primitive grounding (BRepGround) to a promptable setting that also accepts partial selections of target primitives. Compared to BRepGround, the architecture introduces a partial-prompt branch with similarity propagation. The main technical contribution is Discriminative Bootstrapping, which uses the partial branch as a training vehicle: noisy partial prompts built from VLM verification, the model's own predictions, and latent-space hard negatives are fed to the model together with the text. Combined with contrastive and text-alignment losses, this yields a +17.26 F1 gain over the strongest baseline in the original text-only setting.

Critical strengths
  1. Discriminative Bootstrapping turns the new partial-prompt branch into a source of hard-negative supervision, improving text-only F1 by +8.45. Ablations show that each signal source contributes.

  2. The baselines are broad and receive the same pretraining. Replacing BRepGround with PGGM in FutureCAD also improves downstream CAD generation.

Critical weaknesses
  1. The partial and text+partial settings have no baselines.

  2. Figure 3 shows that the model relies on long, fully disambiguating descriptions to select a few distinct primitives that could be picked in a few clicks, and ~70% text-only precision still requires manual correction. The paper does not compare against manual selection or test shorter, realistic instructions.

Questions for authors

Is it possible to add any baseline under the partial and text+partial settings to demonstrate the effectiveness of the model?

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

Related papers

Discussion (0)

Sign in to comment.