acceptodds
Under review as a conference paper at ICLR 2027

DraftCoder:A Benchmark and Model for Dimensioned Multi-View Mechanical Drawing

Abstract

Mechanical engineering drawings convey the design intent of parts through standardized views and complete dimensions, serving as a vital bridge between design and manufacturing. When 3D models are unavailable, generating engineering drawings based on part images and key dimensions still relies on manual labor. To address this, we propose DraftCoder, which uses supervised fine-tuning and reinforcement learning based on execution feedback to convert text-and-image inputs into executable programs, outputting editable multi-view DXF engineering drawings with native dimension annotations. We designed the Hierarchical Drawing Consistency Reward (HDCR) to evaluate geometric measurements and dimensional annotations separately, ensuring that the rewards for dimensions and annotations depend on both view and geometric quality. We also constructed the DimDraft dataset and evaluation benchmark, comprising 5,233 samples and nine categories of part structures, with paired text-image inputs, drawing programs, and reference engineering drawings. In our experiments, the rate of scoreable DXF generation reached 86.26%; the GRPO phase required only 96 training samples, achieving relative improvements of 10.29% and 9.12% over SFT in complete annotation matching rate and layout compliance rate, respectively. The dataset will be made publicly available in the future.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Engineering Drawing, Code Generation;Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Uc6POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DraftCoder, a vision-language model for generating editable, dimensioned multi-view mechanical drawings from part images and textual geometric specifications. The model generates executable Python/ezdxf programs that produce DXF drawings with orthographic views, hidden lines, and native dimension annotations. The authors construct a dataset of 5,233 samples across nine part categories and propose a Hierarchical Drawing-Consistency Reward (HDCR) to optimize drawing quality through reinforcement learning following supervised fine-tuning. Experiments show that DraftCoder substantially outperforms general-purpose vision-language models in DXF generation, view completeness, and annotation quality. Ablation studies further demonstrate improvements from execution-based reinforcement learning, particularly in dimension annotation completeness and layout consistency.

Critical strengths
  1. Practical and well-defined task. The paper addresses the generation of editable multi-view mechanical drawings with native dimension annotations, going beyond simple geometric outlines. This is a meaningful engineering task that requires coordinating orthographic views, geometric dimensions, and annotation placement.
  2. Structured dataset and execution-based learning. The authors construct a dataset of 5,233 samples with paired images, geometric specifications, executable programs, and reference drawings. The proposed hierarchical reward explicitly considers geometry, dimensions, and annotation consistency. The ablation results show improvements over SFT alone, particularly in annotation completeness and layout acceptance, supporting the usefulness of execution-based feedback for this task.
Critical weaknesses
  1. Insufficient experimental comparisons. The main experiments compare DraftCoder primarily with general-purpose vision-language models that have not been fine-tuned for this task. Although the ablation studies demonstrate improvements over SFT, the gains are relatively modest, and some ablated variants outperform the full model on certain metrics. Without stronger task-specific baselines or alternative reward designs, it is difficult to assess the effectiveness of the proposed HDCR.
  2. Limited drawing quality and generalization evidence. Despite achieving a high DXF generation success rate, DraftCoder obtains a Geometry Fidelity score of only 21.20/100 and a Layout Acceptance Rate of 31.03%, indicating that substantial drawing errors remain. Furthermore, the dataset is constructed from a single CAD program source and focuses on nine predefined part categories, with no cross-dataset evaluation. This raises concerns about the method's ability to generalize to more complex, real-world engineering drawings.
Questions for authors
  1. Effectiveness of HDCR. The improvements over SFT are relatively modest, and some reward-ablation variants outperform the full model on certain metrics. Could the authors clarify the specific benefits of the hierarchical reward design compared with a simpler weighted reward, ideally through a direct comparison?
  2. Drawing quality. Despite the high Scorable DXF Rate, Geometry Fidelity is only 21.20/100 and Layout Acceptance Rate is 31.03%. What are the main factors limiting these metrics, and how often do the generated drawings meet practical engineering requirements rather than merely producing valid DXF files?
  3. Generalization. The dataset is derived from GenCAD-Code and focuses on nine predefined part categories. Could the authors provide evidence that DraftCoder generalizes to unseen part structures or independently sourced engineering drawings, beyond the held-out split of the same dataset?
10 comments

Dear Reviewer Uc6P,

Reviewer #Uc6P

We would like to thank the reviewers for their thorough review, their recognition of the research value of this work, and their constructive feedback. Your suggestions regarding the validity of HDCR, the quality of engineering drawings, and the model’s generalizability across data sources have helped us further refine our experimental validation and analysis of the results. We have summarized the relevant weaknesses and issues into the following three areas and address each one in turn, supported by supplementary experiments and analyses.

| Q1. Specific Advantages of a Tiered Reward Design Over a Simple Weighted Reward System(W1&Q1)

A1: The core consideration of HDCR is that the quality of geometry, dimensions, and annotations in engineering drawings is interdependent; they cannot fully compensate for one another. Annotations should be based on correct geometry and dimensions; when the annotated object is incorrect, a well-formatted annotation alone should not receive the same level of reward. Therefore, unlike planar weighting—where each component contributes to the reward independently—HDCR uses the quality of upstream geometry and dimensions to modulate the reward for downstream annotations, thereby encouraging the model to simultaneously improve both the annotated objects and their engineering representation. To verify this design intent, we trained a supplementary model, Flat6, which directly sums the six quality components with fixed weights and does not account for cross-component dependencies, and compared it with the HDCR model described in the paper under a unified testing and evaluation protocol.

Indicators ↑Flat6HDCRImprove
Geometry Fidelity20.9121.20.3
Dimensional Fidelity21.3321.850.53
Annotation Quality41.9442.971.03
Complete Annotation Match Rate38.75%41.16%+2.41
Layout Acceptance Rate28.82%31.03%+2.21

These improvements are reflected in geometric and dimensional fidelity, annotation quality, and the degree of complete annotation matching, and the layout acceptance rate has also increased. This result is consistent with the design intent of HDCR: rewards must not only evaluate individual quality metrics but also reflect the dependency relationship that “the correct object supports the correct annotation.”

| Q2:Factors Affecting Drawing Quality and Practical Application in Engineering(Q2&W2)

Further analysis indicates that the key factors influencing the performance of these metrics are the precise correspondence of local geometry and the consistency across multiple views. Geometry Fidelity simultaneously evaluates feature matching, contour connectivity, curve distances, and arc parameters, while placing particular emphasis on weaker sub-components. Therefore, even if a drawing has established its overall structure and is expressed across multiple views, deviations in the positions of local line segments, hidden edge correspondence, or arc parameters will still significantly impact the overall drawing score. This metric measures the degree to which fine-grained engineering constraints are met, rather than merely whether the part’s overall shape is represented. Layout Acceptance Rate, on the other hand, calculates the proportion of samples that meet a predetermined relative layout threshold, rather than an average score for drawing readability. When the spacing and relative positions between views deviate from the reference layout, the drawing may fail to meet this threshold even if the individual views remain recognizable. Therefore, professional judgment is required to distinguish between projection relationship errors that affect engineering meaning and layout discrepancies that can be resolved simply by adjusting details such as view spacing. These factors precisely highlight the core challenge of engineering drawing generation: not only must parts be represented, but local geometry, dimensional information, and cross-view relationships must also be coordinated simultaneously. This work incorporates these requirements into the learning process using multi-view data with dimensional annotations and a hierarchical consistency reward, and outputs directly editable DXF files. For results where critical engineering information is already correct, engineers can make minor refinements directly, without having to redraw from scratch.

At the same time, we invited three PhDs with backgrounds in mechanical engineering to independently evaluate 100 samples selected through stratified random sampling. The results showed that 57 samples met the specified engineering representation requirements without any modifications, while another 26 required only minor adjustments to drawing details such as text positioning and spacing between views. These samples could be used without modifying any critical geometric, dimensional, or projection information. These results demonstrate that DraftCoder is not only capable of generating valid DXF files but also provides engineering drawings that are ready for use after minimal refinement, thereby reducing the effort required to create drawings from scratch and organize multi-view and dimensional information.

| Q3:Cross-Dataset Generalization on Data from Independent Sources(Q3&W2)

We supplemented our tests with external datasets from CADExpert (100 models) and gNucleus (30 models), whose CAD models were created using sources different from GenCAD-Code. We preserved the geometry of the source models and generated isometric views, geometric descriptions, and engineering drawing reference answers in accordance with CODE’s standardized specifications. All tests directly used the DraftCoder checkpoints from the original paper, without fine-tuning or other parameter updates on the external data, and followed the decoding settings and evaluation methods of the original experiment.

Indicators↑OURSCADExpertgNucleus
Scorable DXF Rate86.26%95.00%63.33%
Geometry Fidelity21.2020.1911.79
Dimensional Fidelity21.8525.2110.90
Complete View Rate70.32%78.00%40.00%
Layout Acceptance Rate31.03%34.00%20.00%

On CADExpert, DraftCoder achieved a DXF scoring rate of 95.00%; its geometric fidelity was close to the industry average, while its dimensional fidelity, complete view rate, and layout acceptance rate all exceeded the results of the CODE test. This indicates that cross-source migration is not only reflected in file executability but also in dimensional representation, multi-view organization, and layout quality—all of which are maintained without the need for training on the target source.

On gNucleus, the tests provide further coverage of the original FreeCAD model sources.The availability rate of DXF files reached 63.33%,the full-view rate and layout acceptance rate were 40.00% and 20.00%, respectively.These results still provide direct evidence that the model can handle parts from another CAD modeling source without the need for fine-tuning with target data.

Reviewer_ttJgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces DraftCoder, a framework that generates executable ezdxf programs from a part image and textual geometric specifications, whose execution produces editable engineering drawings with multi-view geometry and native dimension annotations. Its contributions are three-fold: the DraftCoder framework, a hierarchical drawing-consistency reward called HDCR, and the CODE dataset. Specifically, HDCR is the most important contribution of this work. It compares generated and reference drawings along six dimensions. The authors conducted experiments on CODE with 1,041 evaluation samples and compared open-source and closed-source VLMs, such as Qwen3-VL-8B, GLM-4.6V-Flash, Kimi-VL-A3B, GPT-5.6-SOL, and Claude Opus-5 (whose results are pending). The results show that the proposed DraftCoder outperforms the reported baselines. The authors also conduct ablation studies to demonstrate the effectiveness of HDCR. However, concerns remain regarding the design of HDCR and the experiments.

Critical strengths

This work explores a new task, for which the authors build a new dataset and evaluation metrics, and train a model that outperforms many open-source and closed-source models.

Critical weaknesses
  1. The CODE dataset (also termed DimDraft in the abstract) is presented as a major contribution (Fig. 5), yet the paper shows only two qualitative examples. It is difficult to assess how complex each target drawing is.
  2. Although the paper ablates individual reward groups (Tables 3 and 4), it does not compare HDCR against a flat-aggregation baseline that directly averages the six reward dimensions. In addition, no implementation details are provided on how to compute the reward from the six dimensions.
Questions for authors
  1. The design of the HDCR is relatively complex. What is the motivation behind this design? Why is a hierarchical harmonic mean (Eq. 2 and 3) adopted?
  2. How is each of the six reward dimensions specifically evaluated?
3 comments

Dear Reviewer ttJg,

Reviewer #ttJg

We would like to thank the reviewers for their thorough review, their recognition of the research value of this work, and their constructive feedback regarding the dataset examples and the incentive mechanism. These suggestions have helped us further clarify the complexity of the dataset, the rationale behind the incentive design, and the implementation details. Below, we summarize the relevant weaknesses and issues into three categories and address each one in turn, supported by experimental results and specific explanations.

| Q1. Complexity of Datasets and Examples (W1)

The target complexity of our DimDraft dataset is reflected in three aspects: geometric representation, structural composition, and dimensional annotations. Statistics show that 47.6% of the reference engineering drawings contain hidden lines, there are 953 multi-stage Boolean-constructed parts, and they involve native dimensional annotations such as linear dimensions, diameters, and radii. Therefore, the goal is not simply contour drawing, but rather multi-view engineering drawings that include occlusion representation, composite structure projection, and dimensional positioning.

Specifically, parts with holes or cutouts require distinguishing between visible contours and hidden structures; multi-stage assembled parts require maintaining consistency in local structures and geometric relationships across different views; and dimensional annotations require simultaneously specifying the numerical value, type, associated view, and corresponding geometric location. These requirements collectively constitute the complexity of drafting, from understanding three-dimensional parts to their two-dimensional engineering representation.

In the appendix of this revised draft, we will include representative examples covering all nine categories of parts and various drafting requirements. These examples will showcase the isometric input and the target three-view drawings with dimension annotations, and for each example, we will list the number of drawing elements, hidden lines, and dimension annotations to specifically illustrate the aforementioned structural and annotation requirements.

| Q2:The Design Motivation Behind HDCR and the Effectiveness of Hierarchical Aggregation (Q1&W2)

The core consideration is that the quality of geometry, dimensions, and annotations in engineering drawings is interdependent, rather than consisting of requirements that can be compensated for independently. Complete views and basic structures provide the foundation for geometric consistency; dimensional consistency depends on correct geometric structures; and annotations should be based on correct geometry and dimensions. Therefore, HDCR aggregates upstream quality at each level and uses it to adjust downstream annotation gains. The harmonic mean is adopted because it places greater emphasis on weaker components, thereby suppressing excessive compensation by high-quality components for weak underlying elements, while retaining the feedback generated by improvements in each component. To verify this design expectation, we trained an additional baseline called Flat6, which uses a flat weighted reward: it directly sums the six quality components with fixed weights and does not establish cross-component dependencies. Under a unified testing and evaluation protocol, we compared it with the HDCR model described in the paper; the relevant results are as follows.

Indicators ↑Flat6HDCRImprove
Geometry Fidelity20.9121.20.3
Dimensional Fidelity21.3321.850.53
Annotation Quality41.9442.971.03
Complete Annotation Match Rate38.75%41.16%+2.41
Layout Acceptance Rate28.82%31.03%+2.21

The improvements achieved by HDCR are reflected in the geometric and dimensional quality of annotated objects, as well as in annotation quality and the degree of complete matching; the layout acceptance rate has also increased. These results are consistent with design expectations and provide experimental support for hierarchical dependency modeling: reward design must not only measure individual quality metrics but also reflect the relationships among these quality requirements. By allowing downstream annotation gains to be modulated by upstream quality, HDCR achieves better performance on engineering drawing metrics directly related to this design objective.

| Q3:Specific Calculation Methods for the Six Reward Dimensions (Q2, W2)

The six rewards are computed deterministically from the execution outputs. We parse the generated DXF and establish view, primitive, and annotation correspondences with the reference drawing. Here, S∈{0,1}S\in\{0,1\} serves as a structural-validity gate, while the other applicable components are normalized to [0,1][0,1]. Dimensional and annotation targets used during training are selected according to their coverage in the input text. Specifically:

  1. Structural validity SS: We check whether the DXF is readable, contains non-degenerate geometry, uses supported geometric entity types, and avoids severe annotation duplication or excessive annotation counts. The score is 11 when these checks pass and 00 otherwise.

  2. View coverage VV: After shared-scale alignment, we establish one-to-one correspondences between the generated views and the reference front, top, and side views. The score is the number of successfully assigned semantic views divided by 33. This component measures view completeness; detailed geometric quality is evaluated by GG.

  3. Geometric consistency GG: We separately match visible lines, hidden lines, circles, and arcs. Primitive-matching F1 scores and contour endpoint-connectivity retention form a topology subscore, which is then combined with curve-distance and circle/arc parameter scores through weighted aggregation. The computation accounts for both average quality and weaker subcomponents to assess the recovery of contours and geometric relationships.

  4. Dimensional consistency PP: For text-covered targets with reliable geometric bindings, we measure dimensions directly from the generated geometry rather than reading the displayed annotation values. Supported targets include linear dimensions corresponding to view widths and heights, as well as uniquely bound circle/arc diameters and radii. For an aligned predicted dimension d^i\hat d_i and its reference value did_i, we compute

qi=exp⁡ ⁣(−∣log⁡(d^i/di)∣0.1),P=0.55 mean⁡i(qi)+0.45 min⁡i(qi).q_i=\exp\!\left(-\frac{\left|\log(\hat d_i/d_i)\right|}{0.1}\right), \qquad P=0.55\,\operatorname{mean}_i(q_i)+0.45\,\min_i(q_i).

A target receives a score of 00 if the corresponding predicted view or primitive is missing. This aggregation jointly constrains overall dimensional accuracy and the weakest dimensional target.

  1. Native annotation consistency AA: We perform one-to-one matching between target dimensions and generated native DIMENSION entities, checking annotation presence, type, displayed value, and anchor position. Target-side quality combines the average and worst target-matching quality, and its harmonic mean with prediction-side precision yields AA, accounting for both omissions and redundant annotations. Correct predictions corresponding to reference annotations outside the selected training targets are treated as neutral: they neither contribute reward nor incur a redundancy penalty.

  2. View layout consistency LL: We estimate each view’s displacement relative to the reference using matched line midpoints and circle/arc centers. We then compute pairwise relative-displacement errors between views, normalized by the drawing-sheet diagonal. The mean error eLe_L is converted into L=exp⁡(−eL/0.1)L=\exp(-e_L/0.1). This evaluates relative multi-view layout without penalizing a uniform translation of the entire drawing. Insufficient layout-matching evidence results in a score of 00.

These components are hierarchically aggregated according to Eqs. (2) and (3) to produce the scalar reward used by GRPO.

Dear Reviewer ttJg, thank you again for your careful review and constructive suggestions, which have helped us improve the clarity and completeness of our paper. We have uploaded a revised manuscript incorporating the clarifications and additional evidence discussed in our responses. Specifically, Section 3.3 now explains the motivation for hierarchical harmonic aggregation and the computation of all six reward components; the ablation study includes a direct comparison between HDCR and Flat6 (Table 5); and the supplementary material provides additional qualitative examples illustrating the structural complexity and annotation requirements of DimDraft. We would greatly appreciate your taking another look at these revisions and sharing any remaining questions. If the added explanations and experimental evidence have satisfactorily addressed your concerns, we would be grateful if you would consider updating your assessment and score accordingly. Thank you sincerely for your time and thoughtful consideration.

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

This paper proposes DraftCoder, a framework for generating editable, dimensioned multi-view mechanical drawings from part images and textual geometric specifications. The method combines supervised fine-tuning with GRPO using a Hierarchical Drawing-Consistency Reward (HDCR), which evaluates view completeness, geometric fidelity, physical dimensions, annotation correctness, and layout consistency. The authors also construct CODE, a dataset of 5,233 examples across nine mechanical part categories, with executable drawing programs and reference DXF files. Experiments demonstrate substantial improvements over general-purpose VLMs and moderate gains from execution-grounded reinforcement learning.

Critical strengths
  • Meaningful and practically relevant task formulation. Generating dimensioned multi-view engineering drawings requires coordinating geometric projection, spatial layout, and editable dimension annotations. The paper addresses an important application that extends beyond conventional CAD geometry generation.

  • Valuable dataset and evaluation framework. The CODE dataset provides paired image–text inputs, executable drawing programs, and dimensioned reference drawings. Its construction pipeline preserves geometric and annotation relationships, enabling evaluation of several complementary aspects of mechanical drafting.

  • Promising execution-grounded training results. DraftCoder achieves high DXF generation reliability and improves annotation completeness and layout acceptance over its SFT baseline. The component ablations provide some evidence that dimension-aware feedback benefits drawing generation.

Critical weaknesses
  • Limited methodological novelty and insufficient isolation of the hierarchical reward. DraftCoder builds on existing VLMs, standard SFT, and GRPO, while HDCR primarily combines predefined drawing-quality metrics through heuristic weighting and hierarchical aggregation. Similar execution-grounded reinforcement learning approaches have already been explored in CAD generation. The ablations remove reward components but do not compare against a flat weighted combination of the same metrics, leaving the specific contribution of hierarchical aggregation unclear.

  • The experimental comparisons do not sufficiently establish the contribution of the proposed learning method. The main baselines are general-purpose VLMs without task-specific fine-tuning, making the large performance differences difficult to attribute to HDCR. Compared with the matched SFT baseline, the RL improvements are relatively modest, and some ablated reward variants achieve higher DXF success rates and geometric fidelity. Stronger task-adapted baselines and multi-seed uncertainty estimates are needed to establish the effectiveness of the proposed training strategy.

  • Limited evidence of high-precision drawing quality. Despite an 86.26% Scorable DXF Rate, DraftCoder achieves only 21.20/100 in Geometry Fidelity, 21.85/100 in Dimensional Fidelity, 41.16% in Complete Annotation Match Rate, and 31.03% in Layout Acceptance Rate. These results indicate that producing executable drawings remains substantially easier than satisfying the underlying geometric and dimensional requirements. The paper should better characterize these remaining errors and their implications for practical engineering use.

Questions for authors
  • What improvement does hierarchical aggregation provide over a flat weighted sum of the same six reward components, using identical GRPO settings?

  • Can the authors compare with task-adapted CAD generation baselines under matched training conditions and report multi-seed uncertainty for the RL improvements?

  • What are the dominant causes of the relatively low Geometry and Dimensional Fidelity scores? How frequently do generated drawings satisfy all required dimensions within practical engineering tolerances?

  • How does DraftCoder perform on independently collected mechanical designs or unseen part categories? Can image-only and text-only ablations clarify whether the model learns visual projection reasoning or primarily follows the supplied geometric specifications?

6 comments

Dear Reviewer xuKf,

Reviewer #xuKf

We would like to thank the reviewers for their thorough review, their recognition of the research value of this work, and the specific suggestions they provided. Your comments regarding independent contributions to hierarchical rewards, task-adaptive baselines and training stability, the quality of engineering diagrams, and generalization ability have helped us further refine our experimental validation and methodological descriptions. Below, we have summarized the relevant weaknesses and issues into four categories and will address each one in turn, supported by supplementary experiments and analyses.

| Q1. HDCR’s Methodological Contributions and the Independent Benefits of Hierarchical Aggregation (Q1&W1)

HDCR’s methodological contribution focuses on modeling dependencies among engineering drawing quality requirements. Views and contours provide the foundation for geometric representation; dimensions must align with the actual geometry, and annotations should be associated with the correct geometric objects and dimensions. Therefore, these quality requirements cannot be compensated for independently: when geometry or dimensions contain errors, merely adhering to standard annotation conventions should not yield the same level of reward. HDCR explicitly incorporates these dependencies into execution feedback through harmonic averaging and upstream quality gating, ensuring that downstream dimension and annotation rewards are moderated by upstream quality.

To evaluate this design, we trained Flat6 further: we retained the computational definitions of the same six quality components, added them directly according to predefined weights, and did not establish hierarchical dependencies across components. The table below compares this approach with the HDCR model from the original paper, highlighting metrics directly related to the design objectives.

Indicators ↑Flat6HDCRImprove
Geometry Fidelity20.9121.20.3
Dimensional Fidelity21.3321.850.53
Annotation Quality41.9442.971.03
Complete Annotation Match Rate38.75%41.16%+2.41
Layout Acceptance Rate28.82%31.03%+2.21

As shown in the table, HDCR improved geometric fidelity, dimensional fidelity, and annotation quality by 0.30, 0.53, and 1.03 points, respectively, while the complete annotation match rate increased by 2.41 percentage points. These improvements are reflected in both the geometric and dimensional quality of the annotated objects and the quality and completeness of the final annotations, consistent with the quality correlations emphasized by the hierarchical dependency mechanism. Furthermore, HDCR’s layout acceptance rate increased by 2.21 percentage points. The above results provide experimental support for HDCR’s design philosophy: a reward design must not only measure individual quality metrics but also reflect the dependencies among these quality requirements. By modulating the benefits of downstream annotation based on the quality of upstream geometry and dimensions, HDCR achieved better results on key engineering drawing metrics directly related to this design objective.

| Q2:Task-Adaptive Baseline and multi-seed uncertainty for the RL improvements(Q2&W2)

To verify the reproducibility of the reinforcement learning improvements, we repeated the HDCR reinforcement learning using three different training random seeds, starting from the same SFT checkpoint, while keeping the training configuration, inference settings, and evaluation protocol consistent. The table below reports the mean and standard deviation across seeds for all ten metrics.

指标DraftCoder-SFTDraftCoder-HDCR(均值 ± 标准差)
Scorable DXF Rate(%)83.9686.97 ± 0.59
Structural Validity83.1986.14 ± 0.36
View Coverage76.4379.46 ± 0.77
Geometry Fidelity20.2820.75 ± 0.22
Dimensional Fidelity2121.47 ± 0.72
Annotation Quality40.7641.37 ± 0.94
Layout Consistency51.6353.47 ± 0.71
Complete Annotation Match Rate(%)37.3238.16 ± 1.45
Complete View Rate(%)66.7669.39 ± 1.20
Layout Acceptance Rate(%)28.4329.23 ± 1.33

The results show that HDCR outperformed the SFT baseline on average across all ten metrics. Specifically, the scorable DXF rate and the complete view rate increased by an average of 3.01 and 2.63 percentage points, respectively. In terms of cross-seed consistency, all three seeds showed improvements in metrics such as drawing delivery, view completeness, and geometric fidelity; the standard deviation for Geometric Fidelity was 0.22, while that for the Scorable DXF Rate was 0.59 percentage points. These results indicate that the aforementioned gains from HDCR can be consistently observed across different training random seeds, while also fully quantifying training variations in geometric, dimensional, annotation, and layout quality.

| Q3: Sources of Geometric and Dimensional Errors and Compliance with Engineering Requirements (Q3, W3)

Further analysis indicates that the key factors affecting the performance of these metrics are the precise correspondence of local geometry and the consistent representation of dimensional information across multiple views. Geometry Fidelity simultaneously evaluates feature matching, contour connectivity, curve distances, and arc parameters, while placing emphasis on weaker sub-components. Therefore, even if the drawing has established its overall structure and is represented across multiple views, deviations in the positions of local line segments, hidden edge correspondence, or arc parameters will still significantly impact the overall drawing score. This metric measures the degree to which fine-grained engineering constraints are met, rather than merely whether the part’s overall shape is represented. Dimensional Fidelity, on the other hand, directly examines geometric measurements after alignment to a unified scale and assigns a score based on both the average dimensional quality and the worst-performing dimensional item. Consequently, even if the main contour and some dimensions are correctly represented, deviations in local thickness, hole diameters, or arc parameters—as well as inconsistencies in dimensional scale across different views—will still lower the overall drawing score. This evaluation emphasizes the collective fulfillment of all dimensional requirements; a high score cannot be achieved based solely on partially correct dimensions. These factors precisely highlight the core challenge of engineering drawing generation: not only must the part be represented, but local geometry, dimensional information, and cross-view relationships must also be coordinated simultaneously. This work incorporates these requirements into the learning process using multi-view data with dimension annotations and a hierarchical consistency reward system, and outputs directly editable DXF files. For results where critical engineering information is already correct, engineers can make minor refinements directly, without having to redraw from scratch.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.