- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
CADENA reconstructs editable CAD programs from target meshes by generating one modeling operation at a time. At each step, aligned multi-view renderings of the target and the current reconstruction, together with a three-dimensional hint point, condition a vision-language policy. The proposed operation is appended to the program and executed; geometric matching determines acceptance and stopping. Training combines procedural stepwise supervision with online reinforcement learning based on executed geometry. The paper also introduces CADENA-Bench, containing 3,396 mechanical parts in six families, and GMS, a surface-matching measure using positional and normal agreement. Experiments report improvements over single-pass and prior reconstruction systems. On CADENA-Bench, CADENA-RL obtains GMS 66.95 and a 1.6% invalid rate. The paper acknowledges family-level exceptions, failures of otherwise valid programs, and differences in the input/evaluation protocols used for some external comparisons. (Sections 3–6; Tables 1, 2, 10 and 13.)
Critical strengths
1 Effective operation-level reconstruction with useful internal controls The paper turns geometric execution into feedback for the next modeling decision, with a practical SFT-plus-RL training recipe. The same-backbone, same-generator-source control is particularly useful: on MCB, GMS improves from 57.3 for Single-pass SFT to 63.4 for CADENA-SFT and 73.7 for CADENA-RL, while invalid rates decrease from 18.7% to 12.0% and 0.7%. These comparisons support the overall stepwise design and subsequent RL stage, without establishing the independent effect of every component. The gains are consequential for reconstructing complex mechanical geometry. (Section 5.2; Table 1, p9.)
2 Evaluation resources that distinguish validity from fidelity CADENA-Bench and GMS broaden evaluation beyond simple sketch–extrude shapes and successful program execution. The full benchmark results show CADENA-RL achieving GMS 66.95 versus 54.80 for the strongest external GMS baseline, CADFit, while also reporting invalid rates rather than hiding failures. The metric analysis, category breakdowns, and explicit valid-but-wrong examples make the work more informative and useful to the CAD reconstruction community. GMS remains a geometric proxy, rather than a verifier of construction history. (Sections 4–5; Appendix A.8; Tables 10–12; Figure 5.)
Critical weaknesses
1 The strength of attribution and comparison exceeds the available controls The internal single-pass comparison is useful, but changes in observation, operation granularity, training and inference computation are not separately controlled. Runtime or quality-versus-compute results are not reported, so the practical cost of repeated model calls, execution and rendering remains unclear. External comparisons also have important qualifications: quality means use method-specific valid subsets; CADFit has a fixed optimization budget and a normalization exception; BenchCAD gives CADENA a target mesh while several compared systems consume images; and SOV-CAD is compared using published numbers from a different protocol. The paper discusses many of these limitations, which is welcome, but broad superiority and component-attribution claims should remain limited to the supported settings. This affects the interpretation and practical significance of the reported gains. (Sections 5–6; Appendices A.5–A.6, A.9, A.11 and A.13.)
2 Benchmark provenance is not fully auditable from the manuscript Appendix A.7 describes approximately 60k initial parts, approximately 300 inspected DBSCAN clusters, and one retained representative per group, but does not reconcile those stages with the final 3,396 parts. Noise points or other selection groups may explain the count; the manuscript should make the accounting explicit. The sources also overlap with sources used for RL training, so source names alone cannot establish separation: split identifiers and cross-split near-duplicate checks are needed. This is not evidence of leakage, but it limits independent assessment of generalization. The reported 70% agreement on 254 manually checked family labels further limits interpretation of per-family claims; it does not invalidate label-independent aggregate scores. Auditable selection/split metadata would materially strengthen the benchmark contribution. (Section 3.4; Section 4.1; Appendix A.7; Tables 2 and 10–12.)
Questions for authors
1 Clarification of evaluation and computation Can you provide per-method valid sample counts for each metric, and recompute the main comparisons on a common valid intersection using the existing predictions? Please also give end-to-end timing or quality-versus-compute results for single-pass, stepwise SFT, and stepwise RL inference, including execution/rendering costs. Which factors besides backbone and generator source are controlled? This would distinguish geometric improvement from subset selection and increased computation. (Section 5.1; Appendix A.9.)
2 Clarification of dataset construction and split separation Please reconcile the approximately 60k initial shapes, approximately 300 inspected clusters, retained representatives, and final 3,396 benchmark parts, including the treatment of DBSCAN noise. Can you identify the released sample IDs and document near-duplicate checks against the meshes used for RL training? Please also clarify that family labels affect the composition of per-family means, even though they do not enter the scoring formula or overall aggregate. (Appendix A.7; AI Use Statement, p10.)
3 Clarification needed for implementation and reproducibility Please point to the exact SFT/RL losses, parameter-freezing or adaptation settings, optimizer and learning-rate configuration, and the hint-point handling used in RL. How is the GMS-based stopping score scaled when Section 4.2 defines GMS on 0–100 but inference uses a threshold above 0.99? These details may already be in the anonymized artifacts; identifying the relevant configurations would resolve the manuscript-level ambiguities. (Section 3; Appendices A.1, A.3 and A.8; Reproducibility Statement, p10.)
2 comments
Response to avma, continuation
Separation from RL training. At release, we compared file sizes and then MD5 hashes against the RL training meshes. This removed 108 byte-identical matches, all from MCB. The released benchmark contains 3,396 parts: 1,508 from MCB, 1,602 from TMCAD, and 286 from CADNET.
Effect of family labels
We agree that our statement about the averages was too broad. Family labels determine which parts contribute to each family's statistics, so labeling errors limit the interpretation of per-family results. Policy training, the GMS formula, and the overall aggregate across parts are independent of these labels. We will clarify this in Section 4.1, Appendix A.7, and the AI Use Statement. Agreement with manual labels on the checked subset should inform the interpretation of the family breakdown.
Training details and reproducibility
Section 3 describes next-operation prediction with teacher forcing. Each SFT stage runs for two epochs with a batch size of 64, an initial learning rate of , and a cosine schedule. We will add the following implementation details to the revised version: all model parameters, including the vision encoder, are updated during training (no parameter freezing or adapters); we use the AdamW optimizer with , , and ; the learning rate is linearly warmed up over the first 1000 steps; we apply a weight decay of ; and gradient norms are clipped to .
Section 3 and Appendix A.3 specify an IoU-based geometric reward, zero reward on execution failure, a greedy reference trajectory of up to 10 operations, and 32 sampled continuations per state. We normalize advantages within each group and retain the two candidates with the largest absolute advantages. The effective batch size per update is 48. The main RL run used dr_grpo, clipping with , AdamW, and constant learning rate . The KL penalty was disabled in the main run ().
GMS scale and stopping threshold
The implementation uses GMS on a scale, and the paper and its tables report . The stopping rule uses the internal threshold of 0.99, which corresponds to a reported GMS above 99. For watertight target meshes, the rule uses IoU. For other target meshes, it uses GMS. Both internal scores range from 0 to 1.
Response to avma
Thank you for your positive assessment of our work!
Evaluation comparability, controls, and computational cost
Evaluation protocol. We compared the five external baselines with greedy CADENA-SFT and CADENA-RL runs on parts with valid predictions from all seven methods. Table 1 reports the recomputed metrics and the common sample counts. IoU additionally requires watertight targets and an available value for every method. CADENA-RL leads on all three metrics on each subset.
Table 1. Recomputed metrics on the common valid intersection of all seven methods. CD is the median at 8k sampled points, multiplied by 1000. IoU and GMS are means in percent. The final two rows give the common sample counts for each metric. Bold indicates the best score.
DeepCAD
Fusion360
MCB
Inference cost. To complement the operation counts in Appendix A.6, we measured inference time on 100 MCB parts. With one CPU worker, the median and 90th-percentile latencies per part were 1.88 and 5.30 s for CADENA-RL and 0.96 and 1.18 s for our single-pass model. With eight workers, throughput reached approximately 9,260 and 22,393 parts per hour, respectively. Each method used one H100. CADENA-RL averaged 2.85–2.88 model calls per part, and took 2.42–2.76 times as long (due to the smaller number of output tokens when predicting single operation and KV cache) to process the full sample. In this comparison, inference time scaled roughly linear with the number of model calls.
External comparisons. CADFit uses a separate normalization procedure and a fixed optimization budget. Runs that fail to complete count toward IR, and scores on the remaining subset may overestimate quality. In the BenchCAD comparison, access to the mesh gives CADENA a richer observation and lets it select a prefix using a geometric score. Appendix A.11 describes the differences in inputs and protocols. For SOV-CAD, Appendix A.13 compares published numbers from different protocols. We will qualify our performance claims to reflect these conditions.
CADENA-Bench selection and separation from RL training
We will revise Appendix A.7 to distinguish the clusters used for automatic deduplication from the groups used for manual review. For the three source corpora represented in the release (MCB, TMCAD, and CADNET), the initial pool contained 72,899 meshes. Automatic deduplication left 18,119 candidates. Manual review yielded a preliminary evaluation set of 3,504 parts, and the final check for exact matches with the RL training meshes left 3,396 parts.
Automatic deduplication. We normalized each mesh to the unit cube and encoded it with 3D shape encoder. A transformer trained to predict MCB categories mapped the shape latents to a 64-dimensional embedding, which we L2-normalized before clustering. We clustered the pooled embeddings in three successive passes: DBSCAN (ϵ=0.075,
min_samples=1), HDBSCAN (min_cluster_size=2,cluster_selection_epsilon=10−12), and DBSCAN (ϵ=0.1,min_samples=4). At each pass we kept one representative per cluster and carried forward every point assigned the noise label −1.Manual review. We organized the surviving candidates into 250 groups with BisectingKMeans (
n_clusters=250,n_init=4). One author reviewed all the groups, looking for residual duplicates (including mirrored copies and minor parametric variants), apparent assemblies, and primitives or near-primitives. Each group could contribute several distinct parts. The review applied all exclusion criteria in one pass and yielded the preliminary set of 3,504 parts.