Aggregate analysis

Research Analysis

Cross-paper synthesis of shared research patterns, differences, mainstream directions, and trend evolution.

301 papers

method

New AI methods for CAD generation, reconstruction, editing, retrieval, or reasoning.

39 Methodology quality 79 Recency 43 Reproducibility 25 Topical relevance

Literature review synthesis

Research Lines

Unified multimodal benchmarks and evaluation protocols for CAD program generation

Standardized comparison across text, image, sketch, mesh, point cloud inputs and across geometric, executability, and engineering metrics.

Open edge: Whether benchmark rankings transfer to engineering-grade manufacturability, assemblability, retrieval, editing, and reasoning tasks.

LLM-driven text-to-CAD program synthesis with structure and parameter grounding

Converting text design intent into executable parametric scripts while grounding B-Rep primitives or parameters.

Open edge: Robustness to ambiguous prompts and geometric complexity, and whether gains come from supervision, reinforcement, grounding, or data.

Mesh, image, and point-to-CAD reverse engineering by stepwise or optimization-based reconstruction

Recovering editable parametric operations from geometry while improving validity and geometric fidelity.

Open edge: Scalability to complex topology, behavior under modality shift, and separation of optimization/search from learned priors.

Agentic clarification and specification resolution for text-to-CAD

Reducing invalid or under-specified CAD programs by asking targeted questions before code synthesis.

Open edge: Interaction overhead, generalization beyond curated CadQuery data, and compatibility with non-agentic pipelines.

Shared Direction

CAD is represented as editable parametric programs or construction sequences, making validity, compactness, and geometric fidelity measurable.
Evaluation is trending multimodal and multi-metric, with repeated evidence that model rankings and failure modes change across inputs, geometry complexity, and metrics.
Grounding and alignment mechanisms are needed to connect natural language or visual inputs to executable CAD structure and parameters.
Validity is a first-class concern: executable code can still fail to yield valid geometry or engineering-ready design.
Benchmarks and models increasingly evaluate engineering intent beyond geometric similarity, including functionality, manufacturability, and assemblability.

Key Differences

Generation mechanism: LLM code synthesis, specialized sequence models, and optimization or stepwise reconstruction differ.
Supervision signal: supervised fine-tuning plus reinforcement learning, structure alignment with parameter ranking, and geometric IoU optimization differ.
Evaluation target: geometric/executability/compactness metrics versus engineering-grade manufacturability, functionality, and assemblability have different priorities.
Interaction mode: single-step text-to-CAD versus proactive clarification agents; single-modal versus multimodal inputs.
Model scope: unified multi-task multimodal models versus task-specific mesh-to-CAD or text-to-CAD systems.

Open Questions

How well do current benchmark metrics predict engineering readiness, given the observed failure cascade from executable code to valid geometry to design-intent alignment?
What is the isolated contribution of learned priors versus optimization or search procedures in mesh, point, and image-to-CAD reconstruction?
Can proactive clarification reduce ambiguity without excessive interaction overhead and transfer beyond curated text-to-CadQuery settings?
What evidence gaps remain because limitations, code, datasets, and baseline details are not consistently reported?
How should unified multimodal CAD models be evaluated on retrieval, editing, and reasoning tasks that current evidence does not cover?

CVAILGGRCLSEROCG

250 papers

benchmark

CAD datasets, evaluation tasks, geometric validity metrics, or design benchmarks.

46 Methodology quality 80 Recency 38 Reproducibility 28 Topical relevance

Literature review synthesis

Research Lines

Benchmark task and dataset design

Creates shared CAD evaluation tasks across image, point cloud, mesh, and text inputs for program generation and reverse engineering, with stratified samples and unified task suites.

Open edge: Whether the task distributions and stratified sampling cover real engineering workflows, assembly complexity, and deployment modes beyond the included benchmark families.

Metric and stress-test design

Probes geometric fidelity, executable program validity, compactness, and design-intent criteria such as manufacturability, functionality, and assemblability; tests robustness under modality shifts and ambiguous prompts.

Open edge: Alignment between proxy metrics or VLM-judged rubrics and downstream design decisions, plus stability of model rankings across metrics and difficulty strata.

Comparable baselines and reporting

Enables model comparison through shared evaluation suites, common program validity ratios, and large-scale generation protocols.

Open edge: Whether conclusions remain stable across implementations and compute budgets; several papers have missing code, dataset, artifact, or limitation details, so reproducibility is a verification gap.

Shared Direction

CAD generation and reconstruction are evaluated as executable parametric programs or construction sequences, not only as mesh or point surfaces.
A single geometric score is insufficient; evaluation should combine code validity, geometry quality, program compactness, and engineering design intent.
Specialized and general multimodal models exhibit complementary failure modes, especially under clean versus shifted input modalities.
Stratification by geometric complexity and diversity sampling is valuable for exposing performance degradation.

Key Differences

Evaluation target differs: some work emphasizes reverse-engineering reconstruction accuracy, while others evaluate text-to-CAD design intent, manufacturability, functionality, and assemblability.
Input and task coverage differ: unified multi-task/multi-modal suites contrast with task-specific reverse-engineering benchmarks or text-to-CAD engineering suites.
Methodology differs: direct end-to-end multimodal generation, stepwise feature-by-feature construction, hybrid optimization with geometric feedback, and proactive clarification agents impose different supervision and interaction assumptions.
Metric emphasis differs: volumetric intersection-over-union and Chamfer distance remain prominent in reconstruction work, while validity ratios and rubric-based judges are prominent in engineering-oriented evaluation.

Open Questions

Which evaluation targets and thresholds best predict whether a generated CAD model is usable in real manufacturing and assembly settings?
How stable are model rankings when metrics, input modalities, geometric complexity, and prompting ambiguity change?
How can benchmark sampling and stratification move beyond idealized CAD dataset families to cover real engineering complexity and failure modes?
Are VLM-based rubric judges calibrated and auditable enough to replace or scale human evaluation of design intent?
Which missing artifacts and unreported limitations, including code, datasets, and metric details, are needed to verify the observed benchmark conclusions?

CVAILGGRCLSEROCG

202 papers

system

Runnable CAD systems, toolkits, agents, pipelines, plugins, or design platforms.

42 Methodology quality 79 Recency 39 Reproducibility 27 Topical relevance

Literature review synthesis

Research Lines

Runnable multimodal CAD benchmarks and evaluation stacks

Unify diverse CAD datasets, input modalities, and metrics to expose failure modes such as geometric complexity, modality shift, and ranking instability.

Open edge: How these evaluation stacks transfer to external deployment settings, real engineering constraints, and cross-team reproduction.

LLM-driven parametric CAD program generation

Generate executable CAD scripts from text prompts and ground language references to geometric primitives, improving fidelity and editability.

Open edge: Robustness to ambiguous prompts and coverage beyond script-level validity to real design constraints.

Hierarchical and grounded text-to-CAD reasoning

Decompose CAD construction trees into object-level and part-level reasoning to improve structural consistency and parameter accuracy.

Open edge: Evidence for generalization beyond benchmark datasets and for interaction-level specification resolution.

Mesh, image, and point-cloud to CAD reverse engineering

Recover editable parametric operations from geometric inputs through incremental fitting, optimization, or stepwise geometry comparison.

Open edge: Scalability to complex mechanical parts and robustness across clean, noisy, or photorealistic input modalities.

Shared Direction

All systems treat CAD output as structured and editable: parametric programs, B-rep operations, construction sequences, or executable scripts rather than only static meshes.
There is convergence on multimodal evaluation combining geometric, executability, and program compactness signals to capture more than visual similarity.
LLM or agentic components are increasingly used to translate language or visual inputs into CAD operations or grounding decisions.
Current evaluation reveals a performance gap between producing executable code and producing engineering-ready manufacturable assemblies.
Reproducibility and validity are treated as first-class concerns through code, datasets, metrics, and failure analysis, though missing artifacts remain verification gaps.

Key Differences

Some systems specialize in one modality such as mesh-to-CAD, while others unify text, image, sketch, point cloud, and CAD question answering in one model.
Program representation differs: CadQuery scripts, parametric construction sequences, B-rep primitive grounding, or stepwise feature-by-feature operations.
Supervision differs: supervised fine-tuning with reinforcement learning, ranking-based parameter grounding with perturbations, or agent-generated clarification trajectories.
Evaluation targets differ: geometric fidelity and invalid ratio versus design-intent rubrics such as functionality, manufacturability, and assemblability.
Interaction mode differs: single-pass generation versus proactive clarification before drawing, creating trade-offs in latency, user burden, and robustness.

Open Questions

Can unified CAD systems remain robust under modality shift when specialized models already degrade on noisy, photorealistic, or multi-view inputs?
What evaluation protocol can go beyond executable geometry to reliably measure functionality, manufacturability, and assemblability with valid human or VLM judges?
How should an agent decide when to ask for clarification without high interaction overhead in multi-part or assembly-level CAD tasks?
How well do models trained on benchmark datasets such as Fusion360 and UniCAD transfer to external engineering data and team-specific deployment stacks?
What independent reproduction evidence exists for real-world CAD datasets and pipeline artifacts across different hardware and software environments?

CVAILGGRSECLHCSY

155 papers

application

Applied CAD workflows such as mechanical design, manufacturing, architecture, CAE, CAM, or 3D printing.

45 Methodology quality 78 Recency 39 Reproducibility 28 Topical relevance

Literature review synthesis

Research Lines

Mesh-to-parametric CAD reconstruction

Reverse engineers meshes into editable parametric CAD construction sequences, aiming for geometric fidelity and valid parametric operations.

Open edge: Complex designs still show invalid program ratios, limitation details are often missing, and real scanned or industrial mesh generalization is unproven.

Unified multi-modal and multi-task CAD modeling

Handles heterogeneous CAD tasks such as point-to-CAD reconstruction, text/image-to-CAD generation, and CAD question answering in one framework.

Open edge: Cross-benchmark coverage and reproducibility artifacts remain limited, with little explicit limitation evidence for multimodal task interference.

Visual-language CAD code generation

Generates editable CAD code directly from visual input using fine-tuned vision-language models and large image-code training sets.

Open edge: Generalization to real-world images and unseen operations is only partially demonstrated, leaving out-of-distribution robustness uncertain.

Program evolution and synthetic CAD data generation

Creates complex realistic CAD programs through iterative VLM-guided edits and validation, then augments paired training data for image-to-CAD.

Open edge: No limitation details are provided, and generated complexity may not cover manufacturing constraints, CAE/CAM readiness, or industrial workflows.

Shared Direction

CAD is represented as parametric programs or construction sequences, with editable operations as the target rather than plain shape output.
Geometric consistency and program validity or syntax are shared evaluation concerns across recent methods.
Paired data between non-CAD inputs and programs or code is central to training recent models.
DeepCAD, Fusion 360, and MCB benchmarks recur as common evaluation settings.

Key Differences

Input coverage differs: some methods reconstruct from meshes or point clouds, while others condition on images, sketches, or text.
Optimization and supervision differ: hybrid geometric optimization, offline reinforcement learning, VLM fine-tuning, program evolution, and state-space or GAN-based generation are all present.
Program representation differs between native parametric operation sequences and executable CadQuery Python code.
Evaluation differs between surface-similarity or validity metrics and executable geometric assertions used for diagnostic contrast.
Generalization claims differ: some works report state-of-the-art benchmark scores, while one evidence point shows only partial real-world and unseen-operation generalization.

Open Questions

Do reported mesh and image-to-CAD gains transfer to real-world scans, manufacturing constraints, CAE/CAM workflows, and unseen CAD operations?
What are the specific failure modes for complex or multi-feature parts when limitation fields are often missing in high-ranked evidence?
Can unified multi-modal CAD models retain task-specific accuracy and provide reproducible artifacts beyond reported benchmark wins?
Does design-intent conditioning remain harmful or beneficial across other models, longer prefixes, and diverse CAD domains?
Are current geometric-similarity metrics circular or biased toward surface appearance, and do executable assertions align with engineering validity?

CVLGAIGRSECLCGHC

136 papers

theory

Geometric modeling theory, representation analysis, constraints, or formal design reasoning.

54 Methodology quality 82 Recency 37 Reproducibility 32 Topical relevance

Literature review synthesis

Research Lines

Unified multimodal evaluation of programmatic CAD generation

Measures executability, geometric fidelity, compactness, and image/code QA across clean, noisy, single-view, multi-view, and 3D inputs using large stratified sample sets.

Open edge: Metric rankings and model rankings shift across metrics; specialized models degrade under modality shift; generalization to unseen industrial part families remains limited after fine-tuning and reinforcement learning.

Industrial design-intent evaluation for manufacturable and assemblable CAD

Adds structured design specifications and a code-geometry-design-intent protocol to evaluate functionality, manufacturability, and assemblability of B-Rep assemblies with VLM judges.

Open edge: A failure cascade from executable code to engineering-ready design persists; VLM judge reliability beyond limited human validation and low success on fine-grained engineering criteria remain verification gaps.

LLM and agent-driven program synthesis with structural or parameter grounding

Converts text or image inputs into executable CadQuery or parametric CAD scripts by grounding B-Rep primitives, aligning construction-tree trajectories, or clarifying ambiguous specifications before synthesis.

Open edge: Complex operations such as sweeps and lofts are often simplified to sketch-and-extrude; unseen workflow generalization, interaction overhead, and direct evidence for RL/agentic benefit are unresolved.

Representation-centric generative models for long sequences and editable B-rep surfaces

Uses state-space latent generation or differentiable parametric patches to handle longer parametric CAD sequences and generate editable BReps from point-level supervision.

Open edge: Geometric complexity degrades reconstruction quality; robustness under modality shift, longer-sequence validity, and downstream editability comparisons are not fully established.

Shared Direction

A recurring code-to-geometry-to-design-intent failure cascade appears: models may generate executable code or coarse shapes while missing fine 3D structure and parameter fidelity.
Structured representations and grounding, including construction trees, B-Rep primitives or patches, and program formats, improve validity and fidelity over unstructured generation.
Evaluation is moving beyond geometry and executability toward design intent, manufacturability, functionality, and assemblability.
Input modality shift, geometric complexity, and unseen part families are common boundary conditions that degrade performance.

Key Differences

Representation choices differ: parametric CAD sequences, CadQuery scripts, B-Rep primitive grounding, latent sequence encodings, and differentiable B-rep patches imply different editability and supervision assumptions.
Supervision and optimization differ: point-level self-supervision, curated text-CAD pairs, structured design specifications, supervised fine-tuning, reinforcement learning, and agent-collected clarification trajectories.
Evaluation targets differ: geometric fidelity, executability, and compactness versus industrial code QA and editing versus manufacturing, functionality, and assemblability.
Interaction mode differs: direct end-to-end generation, progressive hierarchical reasoning, and proactive clarification before code synthesis.

Open Questions

Do benchmark gains transfer to unseen industrial part families when current fine-tuning and reinforcement-learning evidence shows limited generalization?
Can models preserve complex parametric operations such as sweeps and lofts rather than substituting simpler sketch-and-extrude patterns?
How should VLM-based design-intent judges be validated for engineering criteria beyond limited human agreement studies?
Can point-level supervision or differentiable surface representations match CAD-specific annotation quality while preserving editability and parametric fidelity?
What formal representation assumptions would align with programmatic CAD generation and yield testable correctness guarantees?

CVAILGGRCLSECEMA

51 papers

survey

Surveys, taxonomies, tutorials, position papers, or roadmap papers for AI4CAD.

14 Methodology quality 76 Recency 51 Reproducibility 10 Topical relevance

Literature review synthesis

Research Lines

B-rep generation and geometric modeling

Generating topologically valid and geometrically complex boundary representation models, which are the standard in industrial CAD.

Open edge: Generalization to other CAD formats and verification on large-scale industrial datasets; the provided evidence does not discuss limitations or failure modes.

3D reconstruction from point clouds

Reconstructing structured 3D building models from unstructured point clouds, potentially using polyhedron-based graph neural networks.

Open edge: No methodology or results are available in the evidence packet, so the specific capabilities and limitations cannot be assessed.

Controllable CAD model generation

Enabling user control over the generation process through hierarchical neural coding, which could support design exploration.

Open edge: The evidence lacks details on the control mechanisms, evaluation metrics, and any constraints on the types of controllable edits.

Dataset creation for manufacturing processes

Providing a multimodal dataset (Slice-100K) for extrusion-based 3D printing, which can facilitate learning-based slicing and toolpath planning.

Open edge: The downstream model performance and the dataset's coverage of manufacturing variability are not discussed in the evidence.

Shared Direction

Deep learning is the primary tool for processing CAD-related geometric data.
Synthetic data plays a crucial role in training because obtaining annotated real CAD or scan data is expensive.
Geometric representation choice (B-rep, mesh, point cloud) is a central design decision that influences model capabilities.

Key Differences

Representation type: some works operate on B-rep structures, while others use meshes or point clouds, leading to different strengths in topological vs. geometric fidelity.
Task granularity: the papers target different stages of the CAD pipeline—generation from scratch, reconstruction from sensors, alignment with scenes, or interaction with CAD UIs—which makes direct comparison difficult.
Evaluation setting: some rely on purely synthetic benchmarks, whereas others incorporate real scanned data; the lack of a common evaluation protocol limits comparability.

Open Questions

Can a unified representation support both generative and discriminative tasks across the CAD pipeline?
What are the minimal annotation requirements to adapt synthetic-trained models to real manufacturing or design scenarios?
How should topological validity be measured and enforced in a learning-based generation system?

GRROCVAI