- 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
-
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.
-
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
-
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.
-
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
- 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?
- Where do the Stage 1 BRep–text pairs come from, and what are the exact dataset splits? Are evaluation parts excluded from alignment pretraining?
- 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.