acceptodds
Under review as a conference paper at ICLR 2027

RealCADBench: Benchmarking parametric CAD modeling via industrial design intents

Abstract

Parametric computer-aided design (CAD) modeling is difficult to evaluate with a single metric. Existing CAD benchmarks often emphasize synthetic or CAD-native settings, limited input modalities, or executability and IoUs alone. We introduce RealCADBench, a benchmark for intent-to-program CAD modeling from real industrial design intents. It contains 12,632 tasks from 19 factory-automation categories and spans text descriptions, 2D engineering drawings, real product pictures, and rendered images for both Part and Assembly modeling. We report results on a 1,770-task evaluation slice: 1,745 Part tasks across four input regimes and RCB-Assm25, a 25-task assembly study used in every reported assembly comparison. Each method generates FreeCAD API Python, which a shared runtime executes to export the 3D model. We evaluate the exported model with executability, Solid IoU, Surface IoU, and a rubric-based visual-semantic identity Judge. Among the nine standalone frontier large models we evaluate, no model leads all four metrics. Across the six frontier-scale large models, executability ranges from to , Solid IoU from to , and Surface IoU from to across the four Part regimes. The highest regime-balanced composite comes from a different model than the leaders on the four component metrics. On RCB-Assm25, Codex with GPT-5.5 improves executability and both IoU metrics over standalone GPT-5.5, but lowers the Judge by pp, leaving GPT-5.5 as the Judge leader. We also observe recurring failure modes, most notably missing fine structures, loss of part identity, and incorrect assembly placement. These results show that execution alone is insufficient to characterize realistic CAD modeling and that frontier models and agents differ substantially across executability, IoUs, and visual-semantic identity.

Primary area:
datasets and benchmarks
Keywords:
AI CAD, Benchmark, Multimodal foundation model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_T7chOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RealCADBench, a benchmark for evaluating parametric CAD modeling from real industrial design intents. It contains 12,632 tasks across 19 factory-automation categories, covering text descriptions, 2D engineering drawings, real product pictures, and rendered images for both part and assembly modeling. The authors propose an evaluation protocol combining executability, Solid IoU, Surface IoU, and a visual-semantic Judge to assess different aspects of CAD generation quality. Experiments on 1,770 tasks with nine frontier models and two coding agents show that strong executability does not necessarily lead to accurate geometry or semantic fidelity, and that model performance varies across input modalities and evaluation metrics.

Critical strengths
  1. Real-world industrial CAD benchmark. RealCADBench covers 12,632 tasks across 19 industrial product categories, with diverse inputs including text, engineering drawings, real product pictures, and rendered images. This provides a valuable testbed for evaluating CAD generation models beyond synthetic or CAD-native datasets.
  2. Multi-dimensional evaluation. The benchmark evaluates executability, geometric accuracy, and visual-semantic identity separately. This is useful because a valid CAD program does not necessarily produce accurate geometry or preserve important product features. The experiments also provide insights into how different models and agents perform across these dimensions.
Critical weaknesses
  1. Concerns about ground-truth reliability. For most non-rendered tasks, the reference 3D models are generated using AI-based reconstruction methods rather than original industrial CAD files. Although the authors apply expert verification, the reconstructed references may contain geometric inaccuracies or assumptions about invisible structures. This raises concerns about the reliability of IoU-based evaluation.
  2. Limited evaluation of parametric CAD modeling. Although the benchmark focuses on parametric CAD modeling, the evaluation only considers the exported STL geometry and does not assess feature histories, parametric constraints, or editability. As a result, the benchmark cannot distinguish a well-structured parametric CAD model from a geometrically similar but non-parametric result, limiting its ability to evaluate the claimed task.
Questions for authors
  1. Ground-truth reliability. For most non-rendered tasks, the reference CAD models are reconstructed using AI-based methods rather than original industrial CAD files. How do the authors ensure that these references are geometrically accurate, especially for hidden or ambiguous structures? Could the authors provide quantitative evidence of their reliability beyond expert screening?
  2. Parametric CAD evaluation. The benchmark evaluates exported STL geometry without directly assessing feature histories, parametric constraints, or editability. Could the authors clarify why these properties are excluded, given the benchmark's focus on parametric CAD modeling?
  3. Judge reliability. The paper validates the visual-semantic Judge mainly through cross-model consistency on six systems. Could the authors provide a direct comparison with independent human expert ratings to establish whether the Judge reliably reflects CAD modeling quality?
3 comments
AuthorsOpenReview ↗

Response to Reviewer T7ch

Thank you for the review. We address reference accuracy, parametric evaluation, and Judge validation separately, using the evidence available in the manuscript.

1. Reference provenance and existing quantitative results.

We maintain that the two reference-construction routes must be assessed separately.

For 373 rendered-image tasks, both the input and reference originate from the same verified STEP asset. Non-rendered tasks use candidate reconstructions generated with Hunyuan-3D, followed by screening and practitioner cross-checks. The selection criteria adapt the Part Judge rubric to a 0-10 scale. These procedures establish the stated provenance and curation process; they do not independently quantify error in hidden or reconstructed geometry.

The appendix separately reports results for the CAD-derived regime. Across six frontier-scale models, mean executability is 0.8123, Solid IoU 0.5379, and Surface IoU 0.2165. These are measurements of model outputs against CAD-derived references, not measurements of the reconstruction-reference pipeline's accuracy. A withheld-CAD audit of that pipeline is not reported.

The Judge compares generated-output renders with input evidence without receiving the reference mesh; text tasks use the corresponding product photograph. Its information differs from that used by IoU, but this does not make it independent ground truth or a test of invisible geometry. Non-rendered IoU should be interpreted as agreement with curated reconstructed references.

2. Geometric reconstruction and parametric quality.

We stand by the geometric measurements within their stated scope.

The required output is an executable FreeCAD program, and the scoring artifact is its exported geometry. Solid IoU measures occupied volume and Surface IoU measures boundaries. Both respond to the geometric consequences of errors in relative dimensions, feature proportions, and spatial structure. Judge Score assesses visible shape, salient features, identity, and functional or assembly plausibility where applicable.

These measurements support comparisons of execution and realized geometry. They do not measure feature-history quality, constraint structure, or successful recomputation after edits. Uniform-scale normalization also removes absolute dimensions. The concern about parametric-model quality therefore identifies an unmeasured capability; we will explicitly restrict the claims to the capabilities assessed by the current protocol. No conclusion about editability or manufacturing readiness follows from STL overlap alone.

3. What the Judge consistency experiment establishes.

We stand by the six-system results as evidence of automated consistency.

Section A.5 reports a consistency check on six BenchCAD systems, separate from RealCADBench construction and prompt development. Two Kimi K2.6 runs differ by 0.90 points on average and preserve 13/15 pairwise system orderings. Each agrees with GPT-5.5 on 14/15 pairs and selects the same top-ranked system.

This experiment uses no human ratings. It cannot answer the request for independent human-Judge agreement, and agreement between automated judges does not rule out shared bias. Expert participation in prompt development and post-checks likewise differs from blinded evaluation on held-out outputs. We do not report a complete independent human-validation study here.

The supported use of the consistency result is to characterize one aspect of the frozen Judge's behavior. Judge scores remain separately reported alongside executability and geometric metrics, with their validation scope stated. A direct human comparison requires documented sample selection, rating protocols, independence from prompt development, and paired scores.

4. Public resource and scope of contribution.

We maintain that the publicly released industrial task resource is a concrete contribution.

RealCADBench-V1.0 is publicly accessible, with inputs, reference assets, and metadata available for inspection. The resource covers industrial text, drawings, photographs, and renders, with part and assembly tasks under a common executable-output protocol. Its existing contribution is this task resource and the separate measurement of execution, geometry, and visual-semantic quality. The reference-audit and human-validation questions remain distinct from the availability of that resource.

Thank you for the response and clarifications. I appreciate the authors' acknowledgment of the current limitations. However, my main concerns about ground-truth reliability and Judge validation remain unresolved. I will maintain my current rate.

AuthorsOpenReview ↗

I think we have largely resolved your main concerns (Let me explain).

First, regarding the reliability of the Judge Score, we conducted a correlation analysis between human expert ratings and Judge Scores. Human ratings were provided on a 0–10 scale, and we discretized the Judge Scores into ten bins for comparison. The resulting Pearson correlation coefficient was approximately 0.71, indicating a positive association between the two assessments, although agreement was not perfect. We also qualitatively compared multiple outputs generated by the same model for identical inputs. In these comparisons, the Judge Score was sensitive to visible structural degradation, assigning lower scores to outputs with more pronounced structural errors.

Second, regarding reference-shape reliability, references outside the rendered-image subset were also assessed by human experts using a dedicated annotation tool. This assessment considered both external geometry and internal structure, where supported by the available input evidence. Image-to-3D models can draw on learned object priors to propose plausible internal structures, and expert assessment helps identify candidates consistent with the inputs. This is particularly relevant to the engineering-drawing subset, where dimensions, sectional views, or other structural details, when present, provide additional constraints on geometry that may not be visible in an ordinary photograph.

We will revise the manuscript to incorporate these clarifications and make the supporting evidence explicit. Specifically, we will describe the human–Judge correlation analysis, including the rating scales, score transformation, sample size, and evaluation protocol, and report the observed Pearson correlation alongside representative qualitative comparisons. We will also expand the reference-construction section to explain the annotation tool, the expert assessment criteria, and how the available input evidence is used to assess external geometry and internal structure. These revisions will distinguish directly verifiable geometric features from plausible but unverified completions, making both the evidence and its scope clearer to readers.

Reviewer_j4AKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RealCADBench, which unifies textual descriptions, 2D engineering drawings, real product photographs, and rendered images of industrial products into tasks requiring the generation of FreeCAD Python programs from input design intents. The full benchmark contains 12,632 tasks across 19 factory-automation product categories, while the paper reports results on 1,745 part tasks and 25 assembly tasks. Evaluation includes execution success, Solid IoU and Surface IoU after pose and uniform-scale alignment, and an input-evidence-based visual-semantic Judge. Profile Average (PA) summarizes these metrics. The authors compare nine standalone models and, on RCB-Assm25, two general-purpose coding agents against their corresponding base models. The main finding is that execution success, geometric reconstruction, and visual-semantic consistency produce different model rankings, and that improvements in execution and geometry from agent systems do not necessarily accompany higher Judge scores.

Critical strengths

Adding surface geometry evaluation beyond solid-volume overlap makes the evaluation more diagnostic. The paper reports both Solid IoU and Surface IoU, computed from filled-voxel and boundary-voxel intersections and unions, respectively (Section 3.3 on page 5 and Equation (2) on page 13). Beyond overall volumetric similarity, this design provides additional assessment of model boundaries, thin-walled structures, and local details. It can help identify cases where the overall shape is close to the target but surface details are poorly recovered. The strength lies in the practical utility of this evaluation design; it does not depend on claiming that Surface IoU is a newly introduced metric, nor does it imply comprehensive validation of engineering details or manufacturing readiness.

Explicitly distinguishing standalone models from agent systems and comparing corresponding configurations has practical value. On the same 25 assembly tasks, the paper compares Codex + GPT-5.5 and Claude Code + Claude Opus 4.8 with their corresponding standalone models, evaluating complete systems that include tool execution, error feedback, and iterative revision (Section 4.1 on page 6 and Table 5 on page 8). This design reveals that different performance metrics can change in different directions. For example, relative to standalone GPT-5.5, Codex increases Solid IoU from 0.2103 to 0.2817 while reducing the Judge score from 76.33 to 69.35. Agent benefits therefore cannot simply be equated with improvements in every aspect of quality. This is a meaningful system-level observation. Figure 8 on page 17 explicitly states that the budgets are not compute-matched, so the results should be interpreted as comparisons between complete configurations.

Critical weaknesses

The expert selection criteria for reference models are insufficiently specified for independent verification, limiting assessment of reference quality and the resulting data distribution. Page 5 reports that approximately 17% of candidate 3D models are retained, with accepted candidates receiving expert scores of at least 8. Appendix A.1 on page 12 (lines 603–615) further describes the involvement of five CAD practitioners and cross-checks. However, the manuscript does not provide a clearly defined scoring range, evaluation dimensions, weights, score anchors, or quality criteria corresponding to "at least 8." It also does not specify how many experts score each candidate, how their scores are aggregated, how disagreements are resolved, or the level of inter-rater agreement. The paper does describe basic checks for corrupted files, duplicate records, ambiguous product identity, non-empty geometry, and exportability (page 4 and page 12, lines 630–636). Thus, the concern is not a complete absence of data-cleaning information, but the lack of transparency in the core quality-scoring process that determines reference acceptance. Non-rendered tasks typically lack paired vendor CAD, so their references are constructed by generative methods and then screened. Whether expert review can reliably distinguish geometric errors from plausible completions of details not constrained by the input is therefore particularly important to the credibility of geometric evaluation. A low retention rate is not inherently a weakness, but the absence of retention and rejection statistics by stage and category makes it difficult to assess how selection changes the category and difficulty distributions. The current procedural description does not sufficiently quantify residual errors in accepted references or explain the effect of selection on benchmark representativeness.

Questions for authors

The following questions seek to clarify reference quality, selection effects, and the sources of low geometric scores. They do not presume that low IoU implies experimental error.

Expert selection criteria: Please provide the complete scoring specification for candidate reference models, including the scale range, evaluation dimensions, weights, score anchors, and the criteria for receiving a score of at least 8. Please also specify the number of raters per instance, score aggregation, disagreement resolution, and inter-rater agreement. Could the authors provide accepted and rejected examples near the threshold to illustrate the actual selection boundary? Please clarify the relationship between this selection score and the Judge rubric used to evaluate outputs from the tested systems, rather than treating the latter as automatically defining the former.

Representativeness after selection and reference reliability: Please clarify the denominator underlying the approximately 17% retention rate, distinguish initial source-data screening from candidate 3D reference selection, and report retention counts and rejection reasons for the main stages and categories. For tasks without paired vendor CAD, how do experts distinguish geometry explicitly constrained by the input from plausible but unverifiable completions? Could the accuracy of the reference-construction and selection pipeline be validated on a representative subset with reliable CAD references or sufficiently specified dimensions, including an assessment of residual errors and their effects on the final evaluation?

Sources of low IoU and release of instance-level results: Some objects in the currently visible examples appear relatively uncomplicated in their external structure, yet Table 5 reports assembly Solid IoU values of 0.1632–0.2817 and Surface IoU values of 0.0864–0.1161. The relationship between apparent complexity and reconstruction difficulty requires more direct evidence. Please provide results stratified by geometric complexity, together with task identifiers, inputs, reference models, generated programs, exported models, instance-level scores, effective sample counts, and run configurations corresponding to the main tables, so that the results can be recomputed. Appendix A.7 already provides some agent renders and diagnostic records. Building on these, could the authors show overlays of aligned predictions and references for representative tasks and conduct appropriate sensitivity analyses to distinguish the contributions of generation errors, reference uncertainty, and alignment or voxelization settings to low scores? If source-data licenses restrict release, please clarify the scope of those restrictions and provide corresponding materials that can legally be made available for verification.

1 comment
AuthorsOpenReview ↗

Response to Reviewer j4AK

We clarify the selection procedure, metric definitions, and scope of available evidence.

1. Selection rubric and scoring records.

We maintain that the stated selection dimensions are concrete, while their validation is a separate issue.

The reference-selection criteria adapt the Part Judge rubric to a 0-10 scale. They cover primary form and proportions, identity-defining features, local-detail completeness, surface refinement, and functional topology/interfaces. The manuscript's threshold of 8 refers to that adapted quality scale.

Section A.1 records five practitioners' participation, not five independent ratings per instance. Historical assignments, aggregation/disagreement records, inter-rater agreement, and threshold-adjacent examples are unavailable. Practitioner involvement alone does not supply these statistics or certify hidden geometry.

Selection and evaluation have different inputs: reference screening concerns candidate references, while the frozen Judge evaluates system outputs against input evidence without seeing the reference mesh or its screening score. This clarifies the evaluation's information flow; a shared rubric does not constitute independent validation.

2. Retention denominator and selection effects.

We maintain that candidate-reference retention and source screening require different denominators.

The approximately 680k-to-366,487 and 200k-to-125,489 counts concern source screening. The approximately 17% in the manuscript concerns generated candidate references passing quality selection, rather than the fraction of original marketplace records retained. Without the underlying historical records, it should remain an approximate rate; it cannot supply exact accepted/rejected counts or category-level retention statistics.

The resource covers 19 industrial categories. Coverage alone does not establish preservation of the source distribution. RCB-Assm25 stratification concerns selection from the accepted 167-task assembly pool; it does not quantify selection effects in constructing that pool. Stage/category counts and rejection reasons would be needed to answer that question quantitatively.

3. Reference accuracy and the role of the CAD-derived subset.

We stand by the CAD-derived subset as a separate evaluation, not as a reference-construction audit.

The 373 rendered-image tasks derive their inputs and references from the same STEP assets, and their results are reported separately. They provide an evaluation regime with direct source-CAD correspondence. Because they bypass the non-rendered reference-construction pipeline, they are not a validation experiment for that pipeline.

A direct audit would reconstruct cases with reliable CAD withheld and then measure accepted-reference errors against it. Such an audit is not reported. Accordingly, non-rendered IoU measures agreement with curated reconstructed references. Judge Score separately evaluates the generated artifact against input evidence; this additional score does not resolve uncertainty in the references themselves.

4. Metric implementation and interpretation of low IoU.

We maintain that the reported IoUs have explicit computational definitions.

Section A.2 specifies signed-PCA alignment, uniform-scale normalization, a shared R=96 grid with 2% padding, undilated boundary occupancy, and a deterministic surface-sampling fallback. Both IoUs reuse the same transform. Solid IoU measures filled-volume overlap; Surface IoU measures boundary overlap. These definitions make clear what the reported numbers measure under the selected protocol.

Qualitative examples show missing structures and incorrect placement, but do not separate generation error, reference uncertainty, and registration/discretization effects. Complexity-stratified analysis, aligned overlays, and sensitivity tests are not reported. Low IoU or visual simplicity alone establishes neither evaluator robustness nor task difficulty. Absolute dimensions and editability are also outside these scores.

5. Availability and reproducibility.

We maintain that public task data and full run reproducibility are distinct deliverables.

RealCADBench-V1.0 is public, with inspectable inputs, references, and metadata. The data card specifies CC BY 4.0; the manuscript also states source-asset license constraints. That label alone is not a source-by-source rights audit.

Executability uses all assigned tasks, whereas each quality term uses evaluator-available artifacts. Reproducing the main tables additionally requires per-metric counts and failure reasons, task/run IDs, generated programs and meshes, and evaluator versions/configurations. The public dataset and auxiliary traces do not establish that this complete run-level package is available. The release makes the underlying task resource inspectable; full table reproducibility remains a distinct requirement.

Reviewer_RBZgOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

RealCADBench introduces 12,632 industrial CAD modeling tasks across 19 categories, covering text, engineering drawings, real product images, rendered images, and assemblies. Models generate FreeCAD programs evaluated using executability, Solid IoU, Surface IoU, and a VLM-based Judge. The paper compares nine foundation models and two coding-agent configurations, reporting results on 1,770 tasks.

Critical strengths
  1. Diverse industrial inputs. The benchmark includes real product images, engineering drawings, text, and rendered images across multiple industrial categories.
  2. Multiple evaluation dimensions. Combining executability, geometric IoUs, and visual-semantic assessment provides a broader evaluation than relying on a single metric.
Critical weaknesses
  1. Questionable ground-truth reliability. Most tasks lack original industrial CAD models. Their reference geometry is reconstructed using AI methods and subsequently screened by experts. While expert review improves quality, it cannot establish accurate hidden geometry, dimensions, or functional interfaces. The benchmark may therefore measure agreement with AI-generated approximations rather than fidelity to actual industrial designs.

  2. Limited evaluation validity and novelty. Despite its large claimed scale, only 1,770 tasks are evaluated, including just 25 assembly cases. Moreover, evaluation focuses on exported STL geometry rather than parametric features, constraints, or editability. The experimental pipeline primarily applies existing models, coding agents, and established metrics, making the contribution beyond dataset collection relatively limited.

Questions for authors
  1. How can the authors establish the geometric correctness of AI-reconstructed references when original manufacturer CAD models are unavailable, particularly for hidden features and functional interfaces?
  2. What fraction of the benchmark has genuine CAD ground truth, and can results be reported separately for original CAD references versus AI-reconstructed references?
  3. Why does the benchmark evaluate only exported STL geometry rather than parametric features, constraints, and editability, despite its focus on parametric CAD modeling?
1 comment
AuthorsOpenReview ↗

Response to Reviewer RBZg

Thank you for the review. We address the reference provenance, measured capabilities, evaluation scale, and resource contribution using the documented protocol and results.

1. Reference provenance and the Judge's inputs.

We maintain that IoU and the Judge use distinct reference information.

For non-rendered tasks, Hunyuan-3D proposes candidate geometry, followed by screening and practitioner cross-checks. The selection criteria adapt the Part Judge rubric to a 0-10 scale and cover form, proportions, defining features, detail, and functional plausibility. These references are curated reconstructions; the screening procedure does not establish the accuracy of unobserved geometry.

IoU compares generated geometry with these references. The Judge instead receives input evidence and rendered views of the generated output, without access to the reference mesh. This is a documented difference between the two measurements, not evidence that the Judge corrects reference errors or provides independent validation of them. For non-rendered inputs, the geometric scores should therefore be interpreted as agreement with curated references.

2. CAD-derived references and separate results.

We stand by the separately reported evaluation on 373 CAD-derived tasks.

The 373 rendered-image tasks derive both input and reference from the same verified STEP asset. They constitute 2.95% of the manuscript's 12,632-task collection and 21.07% of its 1,770 evaluated tasks. Both denominators matter: the first describes the resource composition; the second describes the reported evaluation.

The appendix already reports this regime separately. Across six frontier-scale models, mean executability is 0.8123, Solid IoU 0.5379, and Surface IoU 0.2165. These are results on CAD-derived references. They do not measure the accuracy of the non-rendered reconstruction pipeline, which would require a separate reference audit. The documented CAD-derived count is 373; the number of additional vendor-CAD-paired cases in other regimes has not been audited.

3. What the geometric metrics establish.

We maintain that the current scores support comparisons of realized geometry.

Section 3 requires an executable FreeCAD program and scores the exported geometry. Solid IoU measures occupied-volume overlap, and Surface IoU measures boundary overlap. They capture the geometric consequences of errors in relative dimensions, feature proportions, and spatial structure. Judge Score adds assessment of visible shape, salient features, and product identity, with assembly criteria where applicable.

This protocol evaluates realized geometry across different program implementations. It does not evaluate feature-history quality, constraints, or behavior under edits. Uniform-scale alignment also removes absolute dimensional information. We will make the distinction between geometric reconstruction and parametric-model quality explicit in the claims; the existing measurements support the former.

4. Evaluation scale and information provided by the results.

We maintain that the full resource and the reported evaluation are explicitly distinguished.

The manuscript explicitly separates 12,632 benchmark tasks from the 1,770-task evaluation. The evaluation includes all 568 text, 236 drawing, and 373 rendered-image tasks, plus 568 real-picture parts and 25 assemblies. RCB-Assm25 compares six standalone models and two agents on the same tasks, stratified by category, component count, visibility, and complexity. The assembly results describe this 25-task study rather than establish performance over all 167 assembly tasks.

The component metrics also expose different outcomes. On RCB-Assm25, Codex + GPT-5.5 has Solid IoU 0.2817 versus standalone GPT-5.5's 0.2103, while Judge is 69.35 versus 76.33. These reported configuration-level differences motivate retaining separate geometry and Judge columns. Since quality scores use evaluator-available artifacts and configurations differ, they are not a controlled causal estimate of the harness's effect.

5. Public resource contribution.

We disagree with reducing the contribution to data collection alone.

RealCADBench-V1.0 is publicly accessible, with inputs, reference assets, and metadata available for inspection. The concrete contribution is industrial task coverage across four input regimes and part/assembly tracks, organized around a common executable-output contract and separately reported execution, geometry, and Judge metrics. This gives the community a resource for inspecting cases and evaluating additional systems; it does not depend on introducing a new formula for each component metric.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.