acceptodds
Withdrawn from ICLR 2027

BRepAssemblyLLM: Reasoning over BReps for Executable CAD Assembly

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

CAD assembly requires identifying mating entities between parts and estimating a rigid transformation that yields a valid relative pose. Extending large language models (LLMs) to this task is challenging: native Boundary Representations (BReps) exhibit heterogeneous face–edge graph structures, the candidate mating space grows rapidly with entity count, and assembly inherently couples discrete entity correspondence with continuous pose prediction. We present \method, the first geometry-conditioned LLM framework for pairwise CAD assembly. The framework consists of a topology-aware BRep encoder, a lightweight projector, an LLM, and a CAD assembly executor. Given two CAD parts, the BRep encoder jointly encodes face geometry, edge geometry, and their topological relations, while the projector maps the resulting representations into the LLM embedding space. The LLM then predicts a structured assembly action comprising a pair of mating entities and a rigid transformation matrix specifying their relative pose. Subsequently, the CAD assembly executor executes the predicted action and provides contact and interpenetration feedback, which is used to jointly optimize mating-entity selection and pose prediction via reinforcement learning. On the Fusion 360 Gallery benchmark, \method achieves state-of-the-art performance. These results demonstrate that LLMs can move beyond understanding and generating CAD descriptions or code to directly reason over native BReps and produce executable CAD assembly actions. The code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer-Aided Design, Large Language Models

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_1ETPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper connects a topology-aware BRep encoder to an LLM to predict mating entities and a rigid transformation for two CAD parts. Training combines geometry–text alignment, supervised entity selection, and reinforcement learning using contact and interpenetration feedback from a CAD executor. Experiments on the Fusion 360 Gallery report improvements in both connectivity and pose accuracy.

Critical strengths
  1. The task and output representation are practically useful. I like that predictions refer to native CAD entities and produce executable transformations, rather than just assembly descriptions. This gives the model a concrete geometric task and connects its predictions directly to a CAD executor.

  2. The reported improvements are substantial. Connectivity accuracy increases from JoinABLe’s 79.53% to 89.61%, while pose Chamfer Distance drops from 0.0580 for JoinABLe+Search to 0.0321. These are encouraging gains on both parts of the assembly problem.

Critical weaknesses
  1. The central pose-learning step needs a more concrete explanation. Stage 2 only supervises entity selection, leaving pose learning to Stage 3. However, the contact and overlap rewards are not precisely defined, and it is unclear how predicted matrices are guaranteed to represent valid rigid motions. How the policy initially finds useful pose samples without pose supervision also needs explanation. These details are essential for assessing and reproducing the method, rather than minor implementation choices.

  2. The experiments do not clearly isolate what the LLM and RL contribute. A comparison using the same pretrained BRep encoder with a non-LLM matching/pose predictor would help separate representation improvements from the benefit of the LLM. Also, Table 3 provides no pose results for models without RL. This leaves open whether RL is better than directly supervising pose prediction, rather than simply being the only stage trained to produce poses.

Questions for authors
  1. How are contact and overlap computed, how are rotations parameterized or constrained, and how are invalid outputs handled during RL? Is there any pose initialization or supervision beyond what is described?
  2. Where do the Stage 1 BRep–text pairs come from, and what are the exact dataset splits? Are evaluation parts excluded from alignment pretraining?
  3. How does the method compare with the same encoder using a non-LLM prediction head, and with a pose-supervised variant without RL? These comparisons would materially affect my assessment of the proposed formulation.
Reviewer_pnDdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BRepAssemblyLLM and uses an LLM-based training framework for the executable cad assembly task.

Critical strengths
  1. The final results in the last row of table 1 look good.
  2. Although the ablation study in table 3 reveals some additional problems, the logic itself looks reasonable.
Critical weaknesses
  1. From fig. 1, it seems that this paper does not seriously consider why an llm is the appropriate choice. of course, llms can solve many problems, but why not use a traditional Transformer here? What exactly does the llm bring to this task? What I see is that the authors seem to expect strong performance simply by embedding an llm and then applying training, but I find it difficult to consider this a convincing innovation.
  2. The comparison in table 1 is not fair, cuz different methods use different modalities.
  3. In table 3, most of the improvement seems to come directly from sft. Also, why are CDs only reported for configurations with RL in Table 3?
Questions for authors
  1. Can the authors provide a controlled comparison with a Transformer or similar networks using the same BRep encoder and training setup, to demonstrate what benefit specifically comes from using a llm?
  2. Fig. 1 and Fig. 2 seem to contain a lot of repeated information.
  3. In table 3, the Configuration column and the columns 4,5,6 provide repeated information.
Reviewer_5GQBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BRepAssemblyLLM for pairwise CAD assembly. A topology-aware encoder maps native BRep geometry into LLM tokens, and the model predicts mating entities and a relative rigid transformation. Training combines BRep–text alignment, supervised entity selection, and reinforcement learning with geometric feedback. Experiments on Fusion 360 Gallery evaluate connectivity accuracy and assembly pose prediction.

Critical strengths
  • The model preserves native BRep entity identities while predicting assembly poses. This connects geometric feature learning with explicit mating-entity selection and CAD execution.
  • The results on Fusion 360 Gallery are promising, with improved connectivity accuracy and lower pose Chamfer Distance than the reported baselines. The encoder and RL ablations also show consistent gains.
Critical weaknesses
  1. The transition from entity-selection SFT to pose learning with GRPO is insufficiently explained. Stage II supervises only entity indices, but Stage III introduces transformation prediction without explaining its representation, initialization, or validity constraints. It is also unclear how the initial policy obtains informative geometric rewards when sampled poses have neither contact nor overlap.
  2. The contribution of alignment pretraining is not fully established. Its gains with UV-Net are modest, while an “Ours + SFT only” ablation is missing. This comparison is needed to determine whether alignment benefits the proposed encoder and supports the three-stage training design.
Questions for authors
  1. How are transformations represented, sampled, and constrained during GRPO? Without pose SFT, how does the initial policy obtain informative rewards when sampled poses have neither contact nor overlap? Please clarify whether pose initialization, reward shaping, or sampling constraints are used.
  2. Could you add an “Ours + SFT only” ablation to isolate the contribution of alignment pretraining? Please explain how the encoder is initialized and whether it remains frozen when alignment is omitted.

If the authors can clarify the pose-learning procedure and address the missing alignment ablation, I would consider raising my recommendation to weak acceptance.

Minor suggestions (no rebuttal response needed):

  • Please enlarge the text within the figures.
  • Reporting contact success and interpenetration rates would help assess the practical quality of the predicted CAD assemblies.
  • Please expand the discussion of 3D part assembly in Related Work and clarify how this work relates to existing pose-prediction methods.
  • An autoregressive extension to multi-part assembly, with an analysis of accumulated errors, would be an interesting direction for future work.

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

Related papers

Discussion (0)

Sign in to comment.