acceptodds
Under review as a conference paper at ICLR 2027

Drawing-to-CAD Editing: An End-to-End Benchmark and a Reinforcement Learning Approach

Abstract

Parametric computer-aided design (CAD) editing typically presumes access to underlying source programs. In practical engineering workflows, however, designers frequently need to modify components starting solely from dimensioned 2D engineering drawings and natural language instructions, without access to original CAD models. Existing benchmarks evaluate either drawing-to-CAD reconstruction or text-guided editing from provided source CAD models, leaving this end-to-end task unaddressed. Learning this task poses a fundamental coordination dilemma: two-stage cascaded pipelines suffer from severe error accumulation at inference, direct generation forfeits explicit intermediate geometric feedback, and joint learning with undifferentiated feedback can obscure component-specific quality differences and assign reconstruction credit to editing actions that cannot affect source quality. To tackle these challenges, we introduce CAD-D2E, the first benchmark for drawing-to-CAD editing. CAD-D2E pairs dimensioned orthographic views and textual instructions with executable source and edited programs across 15 edit subtypes and five part families, supporting end-to-end geometric evaluation and intermediate verification. We further propose JORE (Joint Optimization of Reconstruction and Editing), a framework that generates both programs within a unified autoregressive policy. JORE uses CARPO (Causal Advantage Routing for Policy Optimization) to route component-relative advantages according to autoregressive stage dependencies. On CAD-D2E, JORE achieves state-of-the-art performance, surpassing cascaded baselines and standard GRPO by substantial margins, with 78.50% IoU and 87.63% program executability. These results demonstrate the vital role of dependency-aware intermediate supervision in drawing-to-CAD editing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CAD, MLLM, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VW1zOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies generating an edited CAD program from three dimensioned orthographic drawings and a natural-language editing instruction, without a source CAD program at inference. CAD-D2E covers 15 edit subtypes with executable build123d source/target programs. Its evaluation set contains 2,692 editing examples from 940 source parts and is source-disjoint from training. JORE uses a single autoregressive policy to generate a source program followed by an edited program. CARPO combines GDPO-style reward-component normalization with stage-dependent routing: source geometry credit applies only to Source, while final geometry and format credit cover both stages. With a LoRA-adapted Qwen3.5-9B, JORE reports 87.63% final-program executability and 78.50% mean IoU among successful executions, improving on Joint GRPO by 1.96 percentage points. Validation includes direct-generation and cascade comparisons, component/routing ablations, and source-reward weight sensitivity.

Critical strengths
  1. A practically relevant end-to-end task with verifiable supervision. CAD-D2E connects drawing reconstruction and instruction-guided editing without requiring source CAD at inference, allowing reconstruction errors to affect downstream editing. Paired executable source/target programs enable verification of intermediate and final geometry. Source-disjoint evaluation and human inspection of all evaluation examples further strengthen its research value.

  2. Targeted controlled ablations. Joint GRPO, RCN-only, and JORE share initialization, data, verifier, and update budget. Table 2 tests normalization and stage assignment, while Table 4 tests all four cross-stage routing configurations. JORE improves final IoU by 1.56 percentage points over RCN-only and 2.58 points over Stage-local, providing useful empirical support for the routing design under this configuration.

Critical weaknesses

1. Evaluation does not yet establish correctness of the requested edit, limiting interpretation of the benchmark and main result. The task requires a local modification while preserving unrelated geometry, but reports only whole-shape volumetric IoU, global surface Chamfer distance, and executability. Small holes, shallow counterbores, or fillets can leave global similarity high when editing is omitted; Figure 6(e)'s 0.25 mm fillet illustrates this risk. Reference-source and predicted-source no-op comparisons and feature/dimension-level editing-success measures are missing. Appendix A.2 (lines 671–672) further restricts geometric metrics to successful executions, potentially comparing different subsets. Using Table 1's rounded aggregates and assigning failures zero IoU gives approximately 68.79% for JORE, 65.56% for Joint GRPO, and 66.11% for Direct-C1 GRPO; JORE remains ahead, so conditional reporting alone does not overturn its ranking. The more consequential issue is unverified edit sensitivity, compounded by unspecified IoU coordinates, alignment, and scale handling. Independently rescaling predictions and references could conceal absolute-size errors. The protocol should explicitly preserve physical dimensions and verify performance by edit type; 78.50% IoU cannot be interpreted as an editing-completion rate.

2. The value and scope of the additional stage routing require stronger empirical support. RCN explicitly follows GDPO, and Appendix A.1 uses a standard causality/score-function principle. The useful contribution is their application and empirical validation in joint reconstruction/editing, rather than a new fundamental gradient identity. In Table 2, CSA-only scores 76.48% versus GRPO's 76.54%; improvement is observed only in combination with RCN, with JORE gaining 1.56 IoU points over RCN-only. This supports the combination in the evaluated configuration, but independent training repetitions, uncertainty estimates, and gradient/training-dynamics evidence are missing, leaving robustness and the proposed reward-dispersion explanation unverified. The seed count is not reported, so it would be inappropriate to assert that all results are single-seed; Appendix B.1's paired weight-comparison intervals do not quantify training randomness. The starred baselines also lack sufficient details on backbones, training data, build123d/command-sequence or CadQuery adaptations, iterative feedback, and budgets. The weak cascade does not control the editor's GT-source/predicted-source distribution mismatch. These omissions limit attribution of the routing gain and broader superiority claims. Source-disjoint testing within a procedural generator also does not yet establish generalization to real engineering drawings.

Questions for authors
  1. Editing validation: Please report reference-source no-op IoU/CD distributions, scores from retaining each method's predicted source, and feature/key-dimension correctness by edit subtype. Does JORE's advantage remain in completing the requested change while preserving unchanged geometry, particularly for small fillets, holes/counterbores, and bolt-circle edits, which account for approximately 27% of evaluation examples?

  2. Geometry-protocol clarification: Is IoU computed through exact B-rep Boolean volume or approximation? What coordinates, alignment, and scaling are used? Are predictions and references normalized independently, and how are absolute dimensions and dimensional edits preserved? Please report full-test-set IoU with failures assigned zero and geometric tolerance success rates that include failures. A common-success subset is useful as an additional diagnostic, but cannot replace full-set evaluation.

  3. Robustness validation: Are the comparisons among JORE, Joint GRPO, RCN-only, and Stage-local stable across independent training repetitions? Please provide means/dispersion and paired intervals clustered by source part, and specify Appendix B.1's resampling unit and failure handling. Group size G=4 and 741 updates alone do not establish that the differences are noise.

  4. Baseline settings and error attribution: Please specify the starred methods' versions, backbones, training data/budgets, representation adaptations, prompts/API examples, and feedback/turn budgets, particularly the actual configurations of CAD-Editor, IterCAD, and Ortho2CAD. This asks for clarification of reported comparisons, not upgrades to versions released shortly before submission. Are a predicted-source-trained cascade, a GT-source editing upper bound, or comparisons using a fixed reconstruction available to test the explanation for the weak cascade?

  5. Mechanism and theoretical scope: CSA-only shows no observed gain while RCN+CSA does; do training curves or gradient-noise measurements support the proposed explanation? Is Appendix A.1's raw-reward identity intended only as motivation for Equations (5)–(7)? It does not itself establish unbiasedness or guaranteed overall variance reduction for the normalized, length-averaged, clipped update. Please distinguish proved properties from empirical observations.

6 comments

We thank the reviewer for recognizing the practical relevance of the end-to-end task, the verifiable source–target supervision, and the controlled evidence for stage-dependent credit assignment.

1. Validation of Requested CAD Edits (Q1)

No-op comparisons. We compare the reference source, JORE's predicted source, and its final output against the same edited target. The table reports Mean IoU (%) over all test instances, with execution failures assigned zero:

Edit categoryReference-source no-opJORE source no-opJORE final
Feature addition90.8968.5073.54
Feature removal97.7870.2471.26
Geometry modification59.0247.7158.95
Pattern creation79.2254.9673.81
Overall78.9258.7068.79

Final generation improves target agreement in all four categories, with an overall gain of 10.09 percentage points over retaining JORE's predicted source. For bolt-circle edits, which account for 26.97% of test examples, the gain is 18.85 percentage points.

The high reference-source no-op IoU for feature addition and removal reflects the large proportion of unchanged geometry. Larger source–target differences in geometry modification and pattern creation provide more room for improvement in global IoU. Since final generation can also correct reconstruction errors, we further examine the requested changes directly.

Local editing evidence. Our General Response to Reviewers defines IoVGT\mathrm{IoV}_{\mathrm{GT}} and Error volume. The targeted results below address holes, counterbores, fillets, and bolt-circle edits, with chamfers included as another small-feature operation. Correct reconstruction repairs are not counted as errors. Subtype results are per-instance averages over evaluable predictions within each subtype; the overall row pools evaluable predictions across all 15 subtypes. Unevaluable predictions are excluded from these local-metric averages.

Each entry reports IoVGT\mathrm{IoV}_{\mathrm{GT}} (%) ↑ / Error volume (% of GT source) ↓.

Edit subtypeJoint GRPORCN-onlyCSA-onlyJORE
Add hole34.26 / 0.350735.74 / 0.336335.49 / 0.332936.42 / 0.3205
Remove hole (fill)49.47 / 2.514456.38 / 2.155550.18 / 1.976463.65 / 1.7697
Hole → counterbore67.78 / 2.500969.60 / 1.981570.77 / 2.013670.42 / 1.9936
Add chamfer11.23 / 0.013030.21 / 0.025010.13 / 0.001025.12 / 0.0110
Add fillet24.31 / 0.017030.43 / 0.038011.76 / 0.001036.12 / 0.0120
Add bolt circle73.24 / 0.471472.56 / 0.466872.22 / 0.473274.99 / 0.4204
Overall (all 15 subtypes)45.47 / 4.881549.36 / 3.574644.23 / 3.190051.72 / 2.6921

JORE improves both metrics over Joint GRPO in all six targeted subtypes. Across all 15 subtypes, it increases coverage by 6.25 percentage points and reduces Error volume by 44.85%, achieving the best aggregate results among the four configurations.

The ablations show why both metrics matter. For chamfers and fillets, CSA-only's low error volume accompanies substantially lower edit coverage, while RCN-only increases coverage over Joint GRPO but introduces more errors. JORE achieves the best aggregate coverage and error volume, with subtype-specific trade-offs against the individual components. These results support the complementary roles of component normalization and stage assignment.

2. Geometry Evaluation and Failure Handling (Q2)

Geometric evaluation. IoU uses B-Rep Boolean intersection and union volumes, rather than voxelization or point-sampling approximations. Both IoU and CD compare geometries in the original global CAD coordinate system, without geometric alignment or rescaling.

Full-test-set evaluation. With execution failures assigned zero, Mean IoU is 68.79% for JORE, 65.56% for Joint GRPO, and 66.11% for Direct-C1 GRPO.

Accuracy on successful executions. [email protected] and [email protected] report the percentage of each method's successfully executed outputs that meet the corresponding IoU threshold. Execution failures are excluded from these denominators.

Method[email protected] (%) ↑[email protected] (%) ↑
Direct-C1 SFT55.6838.48
Direct-C1 GRPO58.0640.34
Source-GRPO + Edit-GRPO37.1522.55
Joint SFT59.0341.08
Joint GRPO61.3542.05
RCN-only61.9842.87
CSA-only61.0042.64
JORE63.0043.76

JORE achieves the highest accuracy at both thresholds among successful executions.

3. Statistical Reliability and Independent Runs (Q3)

Completed paired comparisons. JORE achieves 78.50% Mean IoU. We report each method's mean over its successful executions in the original test set and the corresponding differences. For paired 95% confidence intervals, we jointly resample source parts across methods with replacement, retain all associated test instances, and recompute each method's mean using its own successful-execution mask within each resample.

ComparatorMean IoU (%) ↑JORE − comparator (pp)Paired 95% CI (pp)
Joint GRPO76.54+1.96[0.78, 3.15]
RCN-only76.94+1.56[0.46, 2.65]
Stage-local75.92+2.58[1.47, 4.43]

All three intervals exclude zero. In particular, the Stage-local comparison supports retaining final-edit feedback for Source tokens beyond conditioning editing on an intermediate reconstruction. These intervals quantify evaluation-sample uncertainty. We will add explicit resampling and failure-handling details to Appendix B.1.

Independent training repetitions. Additional independent runs of JORE, Joint GRPO, RCN-only, and Stage-local with different training seeds are underway under matched training settings and budgets. We will supplement this response with individual-run results and means ± standard deviations.

4. Baseline Implementations and Cascade Controls (Q4)

Implementation settings. Ortho2CAD∗, CAD-Editor∗, and IterCAD∗ are adapted from their official implementations to build123d, using Qwen3.5-9B and fixed task-specific prompts without in-context examples.

Ortho2CAD∗. SFT and Dr. GRPO use 1,028 and 3,161 unique source reconstruction pairs, respectively, with one epoch per stage. Evaluation uses one greedy generation without refinement, allowing 3,072 output tokens within a 6,144-token context.

CAD-Editor∗. The locating and infilling models are trained separately for one epoch on 3,000 editing examples, without RL or selective fine-tuning. Training uses GT sources; evaluation uses predicted sources and predicted masks. Inference consists of two greedy calls without execution feedback or repair, each allowing 3,072 output tokens within a 6,144-token context.

IterCAD∗. Training uses 3,000 multi-turn feedback SFT examples, followed by GSPO with GVPM on 5,932 RL tasks. Evaluation uses one greedy trajectory of up to five turns with execution and rendered-view feedback, allowing 8,192 output tokens per turn within a 32,768-token context.

Source-input controls and joint optimization. To address the source-distribution concern, we train an additional cascade whose editor uses predicted source programs from the SFT stage onward. We also evaluate the original editor with GT sources. The GT-source condition receives additional reference CAD information and serves only as a diagnostic reference. The table reports IoU on successful outputs and full-test-set IoU, with execution failures assigned zero in the latter.

SettingIoU on successful outputs (%) ↑Full-test-set IoU (%) ↑Exec. (%) ↑
GT-source reference (GT CAD provided)87.6781.8093.31
Original Source-GRPO + Edit-GRPO57.6239.7368.95
Predicted-source-trained cascade54.2139.2172.33
JORE78.5068.7987.63

JORE's advantage persists after training the cascade editor on predicted sources: it exceeds the adapted cascade by 24.29 percentage points in IoU on successful outputs, 29.58 percentage points in full-test-set IoU, and 15.30 percentage points in executability. Its advantage therefore holds both among successful outputs and when execution failures are included.

The GT-source reference shows that the original editor achieves high geometric accuracy when supplied with accurate source CAD. In the end-to-end setting, adapting the independently trained editor to predicted sources does not close the gap to JORE. These results support the benefit of jointly optimizing reconstruction and editing, complementing the controlled routing ablations in Table 4, which support directing final-edit feedback to reconstruction.

5. Training Dynamics and Theoretical Motivation (Q5)

Training dynamics. We summarize mean raw final-edit reward (Q1×100Q_1\times100) over identical consecutive update ranges for all four configurations. Updates 5–741 are used, where each method has one contributing run.

MethodUpdates 5–250Updates 251–500Updates 501–741
Joint GRPO69.6369.8170.38
RCN-only69.1069.6169.66
CSA-only68.7669.9369.43
JORE70.0070.7771.34

JORE obtains the highest mean reward in all three intervals, consistent with the gains from combining RCN and CSA in the final evaluation.

Contribution. Section 4.2 explicitly attributes RCN to GDPO. Our contribution is its integration with dependency-aware credit assignment for reconstruction and editing, evaluated through controlled ablations.

Theoretical motivation and scope. The original manuscript already presents the theoretical argument as motivation for stage assignment. Section 4.2 uses “motivates,” and Appendix A.1 explicitly states that the reward-to-stage dependencies in Equation (11) “motivate the assignment of RCN-normalized component advantages in Equation 5.” The proof establishes that a source reward fixed by the completed Source prefix has zero expected score-function contribution to Edit continuations. The manuscript does not claim unbiasedness or overall variance reduction for the complete normalized, length-averaged, and clipped update. The effectiveness of the resulting training design is assessed empirically through the controlled ablations and training dynamics.

Reviewer_itt2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces drawing-to-CAD editing: given three dimensioned orthographic views of a part and a textual editing instruction, with no source CAD model, a model must generate the edited CAD program (build123d). The authors build CAD-D2E, a synthetic benchmark with 15 edit subtypes. Its training sets are a 3,000-example cold-start subset and a 5,932-example RLVR subset. Its 2,692-example test set is fully human-reviewed and shares no source parts with the training data. They also propose JORE, in which a single policy generates the source program and then the edited program in one response. JORE is trained with CARPO, which combines per-component reward normalization (RCN) with causal stage assignment (CSA), so that the source-reconstruction reward is credited only to source tokens. On CAD-D2E, JORE (Qwen3.5-9B with LoRA) reaches 78.50% IoU and 87.63% executability, outperforming general-purpose VLMs, cascaded pipelines, IterCAD, and controlled direct and joint GRPO baselines.

Critical strengths

The task is practically motivated: it reflects a realistic scenario in which only drawings are available and the original CAD model is not. The benchmark is carefully built. The test set is fully human-reviewed, its source parts are disjoint from training, and every example comes with paired executable source and edited programs. The comparison set is broad and includes strong recent systems such as IterCAD. The controlled baselines (Direct-C1 vs. Joint, SFT vs. GRPO) share the same backbone and data. The ablations are well designed; in particular, the routing variants in Table 4 isolate the effect of where credit is assigned from the effect of adding reward components.

Critical weaknesses

There is a noticeable disconnect between the proposed method and the target task. The paper focuses on drawing-to-code editing while introducing CARPO. If CARPO is a general-purpose method, it should be evaluated on other multi-stage tasks, yet the evaluation is restricted to a custom benchmark. If CARPO is specifically tailored for this task, it should demonstrate direct improvements in the editing process itself, but currently there are no key metrics showing such benefits. Furthermore, the evaluation metrics seem insensitive to editing operations because both IoU and CD are global reconstruction metrics.

Questions for authors
  1. In several tables of the paper, the mean CD value is roughly ten times the median CD value, which indicates the existence of significant outliers. The appendix states that the calculation only includes successful data points, implying that a considerable number of successful instances still yield very large errors. However, the authors provide no explanation or analysis for this phenomenon. It would be very helpful if the authors could provide a quantile analysis such as p90 or p95.

  2. According to Table 1, performing direct SFT and reinforcement learning training on the training set already significantly outperforms the baseline. This weakens the argument that the performance gains stem from joint generation combined with CARPO, making the results look more like targeted enhancement on specific data. This inadvertently highlights the value of the authors' data contribution, yet there is no statement regarding dataset open sourcing. Moreover, compared to direct baselines, especially control groups designed to reflect algorithmic innovation such as JORE versus Direct C1 GRPO or Joint GRPO, the improvements are relatively limited. In addition, the paper fails to report the variance or confidence intervals for the main results from Table 1 to Table 4.

  3. According to Table 1, joint generation actually reduces the execution rate in both training stages, where SFT drops from 83.80 to 82.99, and GRPO drops from 86.74 to 85.66, yet the final execution rate increases presumably due to CARPO. However, the paper lacks an analysis explaining how CARPO compensates for this drop in execution rate caused by joint generation. It would be very insightful to understand how CARPO bridges this execution gap, which represents a crucial aspect of the method.

If you can clearly answer my questions, I will be willing to increase my score.

8 comments

We thank the reviewer for the constructive feedback on the task motivation, training design, and empirical evaluation.

1. Connection between CARPO and the Editing Task (Weakness)

Task-specific motivation and scope. CARPO connects source reconstruction to the final editing objective. Source tokens receive both source-quality and final-edit feedback, encouraging accurate reconstruction that supports the requested modification. Edit tokens receive direct credit from final-edit and format feedback, which assess outcomes the Edit stage can influence. Source-quality credit is restricted to Source because subsequent Edit tokens cannot change the source program already generated. Tables 2 and 4 evaluate this coordination through controlled component and routing ablations.

Direct evidence of local editing improvements. Across all 15 edit subtypes, JORE achieves the highest overall edit coverage and the lowest error volume among the four configurations. Relative to Joint GRPO, IoVGT\mathrm{IoV}_{\mathrm{GT}} increases by 6.25 percentage points, while error volume decreases by 44.85%. These results show that greater coverage of the requested changes is accompanied by fewer newly introduced geometric errors. Metric definitions and full results are provided in our General Response to Reviewers.

2. CD Distribution and Tail Errors (Q1)

Executability measures whether a valid shape is produced, while CD measures its geometric agreement with the target. To characterize the CD distribution beyond its mean and median, we additionally report the 90th and 95th percentiles of per-instance normalized CD, using the same successful-execution subsets:

MethodMean CD ↓Median CD ↓P90 CD ↓P95 CD ↓
IterCAD*0.143200.016210.470.87
Direct-C1 GRPO0.146170.016450.370.68
Joint GRPO0.156170.015960.390.73
JORE0.136770.015880.310.64

The quantiles reveal substantial errors among some executable outputs despite the relatively low medians, helping explain the mean–median gap. JORE obtains the lowest P90 and P95, showing that its advantage also appears at these higher-error quantiles.

3. Dataset Contribution, Method Gains, and Statistical Reliability (Q2)

Dataset contribution and public release. CAD-D2E is an explicit contribution of this work. We will publicly release the dataset and state this commitment in the revised manuscript.

Additional benefit beyond the training data. JORE and Joint GRPO share the backbone, cold-start checkpoint, RLVR data, verifier, and training budget. Their comparison therefore evaluates CARPO's additional contribution within the same joint-generation setting. The RCN-only and CSA-only ablations further examine component normalization and stage assignment under the same settings.

Statistical support for the gains. JORE achieves 78.50% Mean IoU. We report each method's mean IoU over its successful executions in the original test set, together with the corresponding differences. To obtain paired 95% confidence intervals, we jointly resample source parts across methods with replacement, retain all associated test examples, and recompute each method's successful-execution mean and the resulting differences in each resample.

ComparatorMean IoU (%) ↑JORE − comparator (pp)Paired 95% CI (pp)
Direct-C1 GRPO76.22+2.28[1.30, 4.23]
Joint GRPO76.54+1.96[0.78, 3.15]
RCN-only76.94+1.56[0.46, 2.65]
CSA-only76.48+2.02[0.84, 3.26]

All four intervals exclude zero. The matched Joint GRPO comparison supports CARPO's additional benefit beyond the shared training data, while the Direct-C1 comparison supports the complete joint-generation framework.

4. Executability (Q3)

Joint generation adds source-prediction and conditioning demands, which may explain the small execution-rate decreases relative to direct generation. Although an invalid final program receives zero final geometry reward, an aggregated advantage can still assign source-quality credit to Edit tokens. CARPO separately normalizes reward components and removes this direct credit path while retaining final-edit feedback for both stages.

Table 2 supports the combined design: RCN-only and CSA-only achieve 85.77% and 85.66% executability, respectively, compared with 85.66% for Joint GRPO. JORE reaches 87.63%, improving by 1.97 pp over Joint GRPO and 0.89 pp over Direct-C1 GRPO. These results are consistent with the benefit of combining component normalization and stage-dependent credit assignment.

With the author–reviewer discussion period drawing to a close, we would like to express our sincere gratitude for your detailed review and constructive suggestions.

To address the concern that global geometric metrics may overlook local editing errors, we introduced complementary local editing metrics across all 15 subtypes. These additional results have also received positive feedback from Reviewer VW1z, who specifically recognized the well-formulated definitions of IoV_GT and Error volume and their value beyond global geometric evaluation.

We have also provided CD quantile analyses (P90/P95), paired confidence intervals for the main comparisons, and further analysis of executability under joint generation. We have additionally clarified CARPO's contribution beyond the shared training setup and confirmed our commitment to publicly release CAD-D2E.

We hope these additional results and clarifications have addressed your concerns. We would greatly appreciate any further thoughts you may have before the discussion period concludes and would be happy to clarify any remaining points.

Thank you again for your time and thoughtful consideration.

With approximately two days remaining in the author–reviewer discussion period, we would like to follow up on our responses and draw your attention to the revised manuscript, available through the PDF link on the submission page. Changes are highlighted in blue.

Appendices D.1–D.2 (Tables 9–14) provide local editing evidence, paired confidence intervals, three-seed results, and training reward summaries, addressing your concerns about CARPO’s editing benefits and the reliability of the gains. Reviewer VW1z has also acknowledged the statistical consistency of the routing gains based on the independent runs and clustered bootstrap analysis. The abstract now explicitly states that CAD-D2E will be publicly released upon acceptance. Our earlier response provides the CD P90/P95 results and executability analysis.

We would greatly appreciate your feedback on whether the updated evidence addresses your concerns, so that we have an opportunity to clarify any remaining points before the discussion closes. Thank you again for your time and consideration.

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

This paper considers how to generate an edited CAD model from engineering drawings and a natural-language request when the original CAD program is unavailable. It introduces the CAD-D2E benchmark and JORE, which generates a reconstruction of the original part followed by an edited program. The proposed training method, CARPO, normalizes reward components separately and controls which stages receive reconstruction and editing feedback. JORE improves geometric similarity and program executability over the reported direct, cascaded, and joint-generation baselines. Ablations examine the contribution of the two training choices.

Critical strengths

The paper provides a useful testbed for studying reconstruction and editing together. The paired source and edited programs allow both stages to be evaluated and provide supervision beyond the final output. The separation of source parts between training and test sets and the manual checks on all test examples is also appreciated. These choices strengthen the benchmark as a basis for comparing methods in the setting studied.

The ablations provide convincing evidence for CARPO within the joint model. The comparison with Joint GRPO uses the same starting checkpoint, data, verifier, and training budget. The authors also examine reward normalization and the assignment of feedback to each stage, including their combination. These comparisons make the improvement more persuasive than the full-system results alone and support the proposed combination of training choices.

Critical weaknesses
  1. The evaluation does not directly establish how reliably the requested edits are completed. The task requires both making the specified change and preserving unaffected geometry, but the reported metrics assess overall shape agreement and program executability. This distinction matters for local changes such as the small-radius fillet, omitting the edit could have little effect on the geometric score.
  2. sketch to CAD editing is a rather niche area compared to sketch to CAD and CAD editing, providing lots of editing tasks in real life scenario are multi turned.
Questions for authors

How do training and test sources differ beyond source identity, particularly in operation structure and parameter ranges? Is there evidence on held-out combinations of geometry and edits, or on drawings produced independently of the dataset generator? This would help assess the breadth of the benchmark and the reported gains.

4 comments

We thank the reviewer for recognizing the value of the CAD-D2E benchmark and our controlled ablation studies, and for the constructive feedback on evaluation and generalization.

1. Editing Evaluation (W1)

The additional local editing evaluation shows that JORE achieves the highest overall edit coverage and the lowest error volume among the four configurations across all 15 edit subtypes. Relative to Joint GRPO, IoVGT\mathrm{IoV}_{\mathrm{GT}} increases by 6.25 percentage points, while error volume decreases by 44.85%. These results show that greater coverage of the requested changes is accompanied by fewer newly introduced geometric errors. Metric definitions and full results are provided in our General Response to Reviewers.

2. Task Relevance and Multi-Turn Editing (W2)

CAD-D2E addresses a practical setting in which dimensioned engineering drawings and an editing request are available, but the source CAD is not. The task therefore requires recovering the source geometry and producing the requested edited design from these inputs. We use one fully specified request to isolate reconstruction–editing coordination under a controlled setting. Multi-turn editing additionally involves maintaining instruction history and cross-turn constraints and is an important extension of this setting.

3. Training–Test Overlap and Generalization (Question)

Structural overlap. We audited source-structure–edit-subtype combinations in our source-disjoint test set against the combined SFT and RL training data. Of the 2,692 test examples, 2,437 (90.53%) have combinations seen during training, while 255 (9.47%) contain combinations absent from both training pools.

Results on the unseen-combination subset are:

MethodMean IoU (%) ↑Mean CD ↓Median CD ↓Exec. (%) ↑
Direct-C1 GRPO61.580.285430.0194574.50
Joint GRPO63.360.267440.0169574.51
RCN-only63.530.252830.0171876.08
CSA-only62.950.267720.0172574.51
JORE64.550.248790.0172076.47

IoU and CD are computed over successful executions.

JORE achieves the highest Mean IoU and executability and the lowest Mean CD on this subset. Its Mean IoU exceeds Joint GRPO and Direct-C1 GRPO by 1.19 and 2.97 percentage points, respectively, showing that the observed advantage extends to unseen structure–edit combinations.

Parameter ranges. Observed ranges largely overlap, with some test values extending beyond the observed training extrema:

ParameterObserved training rangeObserved test range
Thickness increment3.00–8.003.00–10.00
Added-slot width1.10–40.101.40–43.10
Added-rib width6.40–10.886.00–10.88

Independent drawings. Our additional evaluation addresses the reviewer's question about unseen geometry–edit combinations: JORE retains its advantage on source-structure–edit combinations absent from both training pools. Evaluating transfer to independently produced engineering drawings remains an important next step.

With the author–reviewer discussion period drawing to a close, we would like to express our sincere gratitude for your detailed review and constructive suggestions.

To address the shared concern about whether global geometric metrics adequately reflect the requested edits, we introduced complementary local editing metrics across all 15 subtypes. These additional results have also received positive feedback from Reviewer VW1z, who specifically recognized the well-formulated definitions of IoV_GT and Error volume and their value in evaluating local edits beyond global IoU.

We have also provided results on unseen source-structure–edit combinations and a more detailed analysis of the training–test parameter ranges, while clarifying the scope of generalization to independently produced engineering drawings.

We hope these additional results and clarifications have addressed your concerns. We would greatly appreciate any further thoughts you may have before the discussion period concludes and would be happy to clarify any remaining points.

Thank you again for your time and thoughtful consideration.

With approximately two days remaining in the author–reviewer discussion period, we would like to follow up on our responses and draw your attention to the revised manuscript, available through the PDF link on the submission page. Changes are highlighted in blue.

Appendix D.1 (Tables 9–11) includes local editing metrics, no-op comparisons, subtype analyses, and full-test-set results, providing evidence of the requested changes beyond overall shape agreement. Reviewer VW1z, who raised a similar evaluation concern, also recognized the local metric definitions as well-formulated. Appendix E clarifies the scope of generalization to independently produced engineering drawings. Our earlier response provides the structural-overlap, unseen-combination, and parameter-range analyses.

We would greatly appreciate your feedback on whether the updated evidence addresses your concerns, so that we have an opportunity to clarify any remaining points before the discussion closes. Thank you again for your time and consideration.

With less than eight hours remaining in the author–reviewer discussion period, we would like to follow up on our responses concerning editing fidelity, task scope, and train–test overlap. We provided additional local editing metrics and results on unseen structure–edit combinations, alongside clarifications of the task setting and training–test coverage.

The revised manuscript is available through the submission’s PDF link, with changes highlighted in blue. We would greatly appreciate your feedback on whether our responses address your concerns, so that we can clarify any remaining points before the discussion closes.

We believe your response would be of great help to our research and would have a positive impact on the ICLR community.

Thank you again for your time and consideration.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to Reviewers

End-to-End Evaluation and Local Editing Evidence

We thank the reviewers for their constructive feedback. We address the shared concern about whether the observed improvements reflect the requested CAD edits.

End-to-end evaluation. CAD-D2E evaluates the final edited program by its executability and geometric agreement with the reference target. We complement this end-to-end evaluation with two local metrics to analyze the editing behavior of JORE and its controlled ablations, quantifying coverage of the requested changes and newly introduced geometric errors.

Local editing metrics. Let SS and TT denote the ground-truth source and target solids, and S^\widehat{S} and T^\widehat{T} the predicted source and final solids. Let V(⋅)V(\cdot) denote volume. Define the reference additions and removals as A=T∖SA=T\setminus S and R=S∖TR=S\setminus T, and the predicted additions and removals as A^=T^∖S^\widehat{A}=\widehat{T}\setminus\widehat{S} and R^=S^∖T^\widehat{R}=\widehat{S}\setminus\widehat{T}. Both metrics are expressed as percentages.

Intersection over ground-truth edit volume (IoVGT\mathrm{IoV}_{\mathrm{GT}}) measures the fraction of the reference edit volume covered by predicted changes, matching additions and removals by direction:

=100\times\frac{V(A\cap\widehat{A})+V(R\cap\widehat{R})}{V(A)+V(R)}.$$ **Error volume** measures newly introduced incorrect additions/removals, normalized by the ground-truth source volume: $$\text{Error volume} =100\times\frac{V(\widehat{A}\setminus T)+V(\widehat{R}\cap T)}{V(S)}.$$ The two numerator terms respectively measure material newly added outside the reference target and material newly removed from within it. Correct reconstruction repairs are not counted as errors. **Results across all 15 edit subtypes.** We compute each metric per instance and report its average over evaluable predictions across all 15 edit subtypes. Unevaluable predictions are excluded from these local-metric averages. | Method | $\mathrm{IoV}_{\mathrm{GT}}$ (%) ↑ | Error volume (% of GT source) ↓ | | :--- | ---: | ---: | | Joint GRPO | 45.47 | 4.8815 | | RCN-only | 49.36 | 3.5746 | | CSA-only | 44.23 | 3.1900 | | JORE | **51.72** | **2.6921** | JORE achieves the highest overall edit coverage and the lowest error volume among the four configurations. Relative to Joint GRPO, $\mathrm{IoV}_{\mathrm{GT}}$ increases by **6.25 percentage points**, while error volume decreases by **44.85%**. Greater coverage of the requested changes is thus accompanied by fewer newly introduced geometric errors.

Thank you for the careful reassessment and constructive suggestions. For the methods included in our local editing evaluation, we will report the exact evaluable sample rates and include corresponding full-test-set results alongside the conditional metrics.

We will also incorporate the additional evidence into the revised manuscript and clarify the attribution of the gains, theoretical scope, and evaluation limitations as suggested. We will follow up with the completed statistics and the corresponding manuscript changes.

AuthorsOpenReview ↗

Rebuttal Summary

We thank the Area Chairs and reviewers for their careful evaluation and constructive feedback. We appreciate their recognition of the task’s practical relevance, the value of CAD-D2E, and the controlled evidence supporting CARPO.

Our work addresses CAD editing from engineering drawings when the original CAD program is unavailable. CAD-D2E establishes this benchmark, while JORE couples reconstruction and editing with CARPO’s stage-specific credit assignment.

During rebuttal, we strengthened the evidence around three central questions:

1. End-to-end evaluation and local editing — a shared concern across all three reviewers. We complemented final-program evaluation with local editing metrics. Relative to Joint GRPO, JORE achieves 6.25 percentage points higher aggregate edit coverage and 44.85% lower newly introduced error volume on evaluable predictions, providing evidence of improved local editing beyond global IoU.

2. Practical significance and reliability of the gains. Our controlled comparisons use the same backbone, training data, and training budget. Paired confidence intervals and three independent training runs support the reliability of CARPO’s gains. When execution failures are assigned zero, JORE also improves full-test-set IoU by 3.23 percentage points over Joint GRPO.

3. Task scope and generalization. We clarified the task scope and disclosed train–test structure–edit overlap and parameter ranges. On unseen structure–edit combinations, JORE achieves 64.55% mean IoU on successful executions, exceeding Joint GRPO and Direct-C1 GRPO by 1.19 and 2.97 percentage points, respectively.

We also provided no-op and predicted-source cascade controls for Reviewer VW1z and clarified gain attribution, theoretical scope, and limitations.

Reviewer VW1z noted that the new results substantially improved the work’s empirical credibility and diagnostic value, and indicated that they were leaning toward raising their score conditional on the requested revisions, while also awaiting the other reviewers’ feedback. We have incorporated the requested findings, evaluation details, and limitations into the revised manuscript and are awaiting Reviewer VW1z’s updated assessment.

Reviewer itt2 recognized the carefully constructed benchmark and controlled experimental comparisons, and appreciated the additional experiments and clarifications. We also directly addressed Reviewer wmRt’s concerns through detailed responses and additional evidence, but regret that we have not yet received further feedback despite our follow-up.

The revised manuscript is available through the submission’s PDF link, with changes highlighted in blue. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.