acceptodds
Under review as a conference paper at ICLR 2027

CADENA: Stepwise CAD Reverse Engineering

21 views

Abstract

Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the intermediate geometry. In contrast, human engineers build a part feature by feature, checking after each operation what remains to be modeled. We introduce CADENA (Spanish for "chain"), a model that reconstructs a 3D mesh as a parametric CAD program, growing its sequence of operations one at a time and comparing the target with the currently predicted geometry at every step. We also address the lack of benchmarks for evaluating reverse-engineering methods on mechanical parts, introducing CADENA-Bench, a benchmark that measures performance across categories of mechanical parts. CADENA outperforms prior methods on CADENA-Bench and on the DeepCAD, Fusion 360, MCB, and BenchCAD datasets.

Primary area:
generative models
Keywords:
Computer Vision, 3DCV, CAD

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_avmaOpenReview ↗
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
AuthorsOpenReview ↗

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

MethodCD ↓IoU ↑GMS ↑
CAD-Recode0.171892.5693.50
cadrille0.169291.9896.03
CADReasoner0.162692.8095.84
CADEvolve0.160194.0796.33
CADFit0.174092.6291.19
CADENA-SFT0.160793.5796.18
CADENA-RL0.150997.0997.78
N (CD, GMS)5,085—5,085
N (IoU)—4,813—

Fusion360

MethodCD ↓IoU ↑GMS ↑
CAD-Recode0.163789.3587.84
cadrille0.160789.0090.53
CADReasoner0.145889.2690.43
CADEvolve0.153391.4391.01
CADFit0.158090.9887.46
CADENA-SFT0.137991.7493.00
CADENA-RL0.126695.9595.26
N (CD, GMS)999—999
N (IoU)—923—

MCB

MethodCD ↓IoU ↑GMS ↑
CAD-Recode0.382776.0061.06
cadrille0.448073.2663.78
CADReasoner0.886673.3958.82
CADEvolve0.336480.7866.67
CADFit0.292381.9166.32
CADENA-SFT0.207784.9572.09
CADENA-RL0.136893.4881.28
N (CD, GMS)1,593—1,593
N (IoU)—849—

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 L2L_2-normalized before clustering. We clustered the pooled embeddings in three successive passes: DBSCAN (ϵ=0.075\epsilon=0.075, min_samples=1), HDBSCAN (min_cluster_size=2, cluster_selection_epsilon=10−12=10^{-12}), and DBSCAN (ϵ=0.1\epsilon=0.1, min_samples=4). At each pass we kept one representative per cluster and carried forward every point assigned the noise label −1-1.

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.

AuthorsOpenReview ↗

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 1.2×10−41.2\times10^{-4}, 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 β1=0.9\beta_1=0.9, β2=0.999\beta_2=0.999, and ϵ=10−8\epsilon=10^{-8}; the learning rate is linearly warmed up over the first 1000 steps; we apply a weight decay of 0.010.01; and gradient norms are clipped to 1.01.0.

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 ϵ=0.1\epsilon=0.1, AdamW, and constant learning rate 3×10−53\times10^{-5}. The KL penalty was disabled in the main run (β=0\beta=0).

GMS scale and stopping threshold

The implementation uses GMS on a [0,1][0,1] scale, and the paper and its tables report 100×GMS100\times\mathrm{GMS}. 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.

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

The paper introduces CADENA, a system for reconstructing editable parametric CAD programs from input meshes. Rather than generating an entire program in one pass, CADENA predicts and executes one operation at a time, conditioning on multiview renderings of the target and current reconstruction, together with a spatial hint. The model is trained on synthetic stepwise supervision and subsequently refined through online reinforcement learning using geometric IoU rewards. The paper also introduces CADENA-Bench, comprising 3,396 mechanical parts across six families, and GMS, a surface-matching metric incorporating positional and normal agreement. Evaluation on DeepCAD, Fusion360, MCB, and CADENA-Bench shows improved reconstruction accuracy and generally low invalid-output rates, with larger gains on complex mechanical parts. Comparisons against a single-pass model trained from the same generator, along with reinforcement-learning and inference-budget ablations, support the benefits of the overall approach. Additional BenchCAD results are reported, although differences in input access limit direct comparison. The evaluation primarily establishes geometric reconstruction quality rather than recovery of the original construction history or design intent.

Critical strengths
  • The results on complex mechanical parts are convincing. On MCB, CADENA improves GMS from the best baseline’s 60.1 to 73.7, with only 0.7% invalid outputs. IoU and Chamfer distance also improve, so the gains are not limited to the authors’ proposed metric. These results are the main reason I lean toward acceptance.

  • The paper includes a useful comparison with single-pass generation. Using the same backbone and procedural data generator, stepwise SFT improves MCB GMS from 57.3 to 63.4 and reduces invalid outputs from 18.7% to 12.0%. This supports the decision to build the model one operation at a time, even though it does not tell us which part of the stepwise setup matters most.

Critical weaknesses
  • The core idea is not new. Earlier work already uses execution and visual feedback to revise CAD programs. Query2CAD [1] uses visual feedback in a refinement loop, while CADReasoner [2] uses geometric discrepancies to iteratively improve programs for the same reverse-engineering task. The concurrent SOV-CAD [3] is even closer: it uses views of the target and the partially constructed model to guide stepwise modeling. CADENA differs in its geometry-only state representation, operation-level appending, and training procedure, but I see these as improvements to an existing approach rather than a new approach.

  • The single-pass SFT baseline is a useful control: it uses the same backbone and procedural generator, and the results support the stepwise setup, particularly on MCB. However, this comparison is limited to SFT. A single-pass baseline with the same RL setup would help show how much of the final model’s advantage comes from stepwise generation. The current experiments also leave open the separate contributions of intermediate geometric feedback and spatial hints, as well as the added inference cost.

[1] Badagabettu et al. Query2CAD: Generating CAD Models Using Natural Language Queries. 2024.

[2] Kabisov et al. CADReasoner: Iterative Program Editing for CAD Reverse Engineering. 2026.

[3] Feng et al. SOV-CAD: Stepwise Orthographic Views Guided CAD Modeling Sequence Reconstruction. 2026.

Questions for authors

See weaknesses

Reviewer_W4FZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CADENA, a stepwise CAD reverse-engineering framework that reconstructs editable parametric CAD programs from 3D meshes. Unlike conventional autoregressive approaches that generate entire programs in a single pass, CADENA generates one operation at a time, using visual feedback from the target geometry and the current reconstruction to guide subsequent operations. The model is trained through supervised fine-tuning on procedurally generated construction sequences, followed by online reinforcement learning using execution-based geometric rewards. The authors also introduce CADENA-Bench and Generalized Match Score (GMS). Experiments demonstrate substantial improvements across multiple benchmarks.

Critical strengths
  1. Strong empirical performance. CADENA achieves substantial improvements over competitive baselines across multiple benchmarks, particularly on complex mechanical parts. The improvements in geometric fidelity and execution validity are convincing.
  2. Comprehensive experimental evaluation. The paper compares against several recent methods using their released checkpoints and provides meaningful ablations separating the benefits of stepwise generation, reinforcement learning, and inference-time strategies.
  3. Table 1 shows an interesting new discovery that the performance on DeepCAD is saturated, with relatively small differences among leading methods. In contrast, performance gaps become larger on more challenging benchmarks such as Fusion360 and MCB.
  4. Valuable benchmark/metric contribution. CADENA-Bench extends evaluation beyond relatively simple CAD reconstruction datasets to diverse mechanical parts. The proposed GMS metric also provides a useful complementary perspective to conventional IoU and Chamfer Distance by incorporating surface-normal consistency. I believe it will be beneficial to the community as the CAD evaluation needs a better way to do automated evaluation.
Critical weaknesses
  1. I am not sure if the authors can claim full credit for stepwise SFT+RL tuning. It looks similar to CADFusion. A clearer discussion between the two works in terms of the training paradigm is appreciated.
  2. The method relies on a CadQuery-based operation vocabulary, which limits the range of supported construction histories. Some generated programs may not transfer faithfully to industrial CAD software.
  3. Unlike single-pass approaches, CADENA requires executing, rendering, and evaluating intermediate geometry after every generated operation, introducing additional inference overhead that may limit scalability.
Questions for authors

How many round of passing does the method typically needed for generating one shape?

Compare with other SFT+RL papers, CADENA is more compute-heavy, right? Can the author provide a more detailed computation on how much more training is needed to approach the same performance level?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We thank all reviewers for their feedback. We have uploaded a revised PDF. All new or changed text is shown in blue. New tables are numbered R1–R2 so that existing table and section numbers are unchanged. Reviewer W4FZ

  • Relationship to CADFusion. The first contribution in the Introduction (p. 2) now explains how CADENA's training differs from SFT-plus-RL pipelines such as CADFusion. Both of its stages operate on single operations: SFT predicts the next operation from the executed program prefix, and RL scores groups of alternative one-step continuations branched from the policy's own trajectory.
  • Training and inference cost.
    • The logged RL cost (48.8 h on four H100 GPUs, about 195 GPU-hours) is in Appendix A.3 (p. 18).
    • Measured inference latency, throughput and model calls per part for CADENA-RL vs. the single-pass model are in a new paragraph, "Inference cost", in Appendix A.6 (p. 22).

Reviewer WuyP

  • Query2CAD. Added to Related Work, Section 2 (p. 3).
  • Scope of the single-pass control. Section 5.2 (p. 9) now states that this control isolates the stepwise setup only at the SFT stage. It also states that the comparison of CADENA-SFT with CADENA-RL measures the effect of RL within that setup. Attributing how much of the final model's advantage comes from stepwise generation would require a single-pass model trained with the same RL, which we do not report.
  • Hint-point ablation.
    • New paragraph "Effect of the hint point" and Table R1 in Appendix A.1 (p. 16), referenced from Section 5.2.
    • The training hint frequency (80% of SFT examples) is now stated in Section 3.4 (p. 5) and in Appendix A.1.

Reviewer avma

  • Qualified performance claims and common valid intersection.
    • Section 5.1 (p. 8) now notes the conditions of the external comparisons: CADFit's normalisation and fixed optimisation budget, mesh input on BenchCAD, and published SOV-CAD numbers from a different protocol. The first contribution is qualified accordingly (p. 2).
    • A new table (Table R2) in Appendix A.9 (p. 25) recomputes all metrics on the parts that are valid for all seven methods. CADENA-RL leads on every metric on each dataset.
  • Benchmark provenance and split separation.
    • Section 4.1 (p. 7) and Appendix A.7 (p. 23) now account explicitly for every stage of the benchmark construction: 72,899 initial meshes, then 18,119 after automatic deduplication, then 3,504 after manual review, and 3,396 after removing exact matches with the RL training meshes.
    • Appendix A.7 now separates the three-pass embedding clustering used for deduplication (with its parameters) from the 250 BisectingKMeans groups used for manual review.
    • It also describes the MD5 check against the RL training meshes (108 byte-identical matches removed, all from MCB) and gives the per-source composition: 1,508 MCB, 1,602 TMCAD and 286 CADNET.
  • Family labels. Section 4.1 (p. 7), Appendix A.7 and the AI Use Statement (p. 10) now state that family labels determine which parts contribute to per-family statistics. Policy training, the GMS formula and the aggregate over all parts do not depend on them.
  • Implementation details. A new paragraph, "Optimisation settings", in Appendix A.3 (p. 18) gives:
    • the SFT settings: full-parameter updates including the vision encoder; AdamW with β1 = 0.9, β2 = 0.999, ε = 1e-8; 1000-step linear warm-up; weight decay 0.01; gradient clipping at 1.0;
    • the RL settings: dr_grpo loss, clipping ε = 0.1, constant learning rate 3e-5, group-normalised advantages, effective batch size 48, KL penalty disabled.
  • GMS scale and stopping threshold. Section 3.5 (p. 5) and Appendix A.3 (p. 18) now state that scores are on a [0, 1] scale internally and that tables report 100×. The 0.99 stopping threshold therefore corresponds to a reported IoU or GMS above 99.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.