Aggregate analysis

Research Analysis

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

284 papers

method

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

37 Methodology quality 78 Recency 43 Reproducibility 25 Topical relevance

Literature review synthesis

Research Lines

Unified multimodal CAD benchmarks and evaluation protocols (e.g., CADBench, MUSE, UniCAD)

Systematic comparison of CAD program generation and reconstruction across input modalities and metrics, including geometric fidelity, executability, and engineering intent.

Open edge: Whether benchmark metrics and VLM judges represent real manufacturability and transfer beyond the selected dataset families; some works lack code or artifacts for verification.

LLM-driven text-to-CAD program generation with grounding and clarification (e.g., FutureCAD, HierCAD, ProCAD)

Translating natural language into executable parametric CAD scripts or B-Rep primitives through fine-tuning, structural alignment, parameter grounding, reinforcement learning, and proactive specification clarification.

Open edge: Robustness to ambiguous specifications and unseen part complexity; how much gain comes from model capability, training data, grounding, or clarification is not fully isolated.

Mesh/image-to-CAD reconstruction and reverse engineering (e.g., CADFit, CADENA)

Recovering editable parametric CAD operation sequences from 3D meshes or images using stepwise construction or hybrid optimization to improve geometric accuracy and program validity.

Open edge: Reliability under high geometric complexity, diverse operation types, and non-ideal input modalities remains insufficiently evidenced.

Universal multimodal multi-task CAD models (e.g., UniCAD-MLLM)

Handling heterogeneous CAD tasks such as reconstruction, generation, and question answering in one model across text, images, sketches, and point clouds.

Open edge: Per-task failure boundaries, cross-task interference, and whether unified models match specialized models under domain shift remain unclear.

Shared Direction

CAD is generally treated as structured programs or construction sequences rather than unstructured geometry alone.
Most methods use parametric, B-Rep, or CadQuery-style representations and evaluate at least one validity or executability measure.
Strong geometric metrics do not guarantee engineering-ready CAD; a gap remains between executable code, valid geometry, and manufacturable design.
Specialized models can outperform general vision-language models under idealized geometric inputs, while generalization under shift is limited.

Key Differences

Evaluation target differs: geometric reconstruction fidelity, program executability and validity, or manufacturability/functionality/assemblability; model rankings change across metrics.
Input and supervision differ: text-only, image, mesh, sketch, point cloud, or multimodal; specialized mesh-to-CAD models differ from general LLM/VLM generators.
Representation and generation mechanism differ: direct script generation, B-Rep primitive grounding, stepwise operation fitting, or optimization over structured programs.
Interaction mode differs: single-shot generation versus proactive clarification before synthesis.
System scope differs: unified multi-task models versus task-specific reconstruction or generation pipelines.

Open Questions

How do methods maintain robustness under modality shift and increasing B-Rep face count? Evidence from CADBench indicates specialized models are brittle under modality shift.
How much of reported performance gain comes from representation, training data, parameter grounding, proactive clarification, or evaluation protocol design?
Can VLM-judge-based design-intent metrics replace human engineering assessment for manufacturability, functionality, and assemblability at scale?
Several papers lack visible code, dataset, or limitation details, leaving reproducibility and failure boundaries as verification gaps.

CVAILGGRCLSEROCG

234 papers

benchmark

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

45 Methodology quality 79 Recency 39 Reproducibility 28 Topical relevance

Literature review synthesis

Research Lines

Benchmark task and dataset design

Creates shared CAD evaluation suites spanning text, image, mesh, point-cloud, and sketch inputs with multi-task outputs such as program generation, reverse engineering, and question answering.

Open edge: Coverage of real design workflows, complex assemblies, manufacturing constraints, noisy inputs, and distribution shift is not established; several datasets and artifacts lack visibility or limitations.

Metric and stress-test design

Probes beyond aggregate accuracy through executability, validity, geometric fidelity, compactness, design-intent rubrics, and staged checks for engineering readiness.

Open edge: Alignment between proxy metrics and downstream engineering decisions is unproven; model rankings change across metrics, and VLM-judge reliability or geometric metric sensitivity is not fully verified.

Comparable baselines and reporting

Supports auditability by reusing shared benchmarks and comparing task-specific models, LLM/VLM models, hybrid optimization, and agentic methods under common metrics.

Open edge: Stability across implementations, compute budgets, and reproductions is unclear; multiple top papers have missing limitation details or limited code/artifact evidence.

Shared Direction

CAD evaluation requires multiple signals rather than a single aggregate score, including validity, geometric error, and design-specific criteria.
Benchmarks increasingly reuse existing CAD dataset families and stratify by complexity to enable controlled comparison across model types.
There is a recurring cascade from executable code to geometrically valid output to engineering-ready design, with current models showing different bottlenecks.
General-purpose language or vision models and CAD-specialized models are compared under common tasks, revealing specialization-performance tradeoffs under clean and shifted inputs.

Key Differences

Evaluation target differs: some works emphasize low-level geometry such as IoU, Chamfer distance, and invalid ratio, while others assess manufacturability, functionality, assemblability, or broader CAD question answering.
Input-output focus differs: text/image-to-CAD generation, mesh/point-to-CAD reconstruction, and unified multimodal multi-task CAD are not the same problem setting.
Interaction mode differs: one-shot program generation is contrasted with proactive clarification and stepwise construction, which trade interaction overhead against robustness.
Supervision and optimization differ: supervised fine-tuning with reinforcement learning, ranking-based parameter grounding, hybrid geometric optimization, and VLM-judge feedback are used separately or in combination.
Representation choice differs: some methods generate parametric construction sequences or CadQuery scripts, while others reason over B-Rep primitives and assemblies.

Open Questions

How well do current geometric and design-intent metrics predict downstream usefulness in real CAD workflows, given that model rankings change across metrics?
What benchmark task distributions are needed to cover noisy, ambiguous, or cross-modal inputs without making evaluation brittle or overly curated?
How stable are reported benchmark conclusions across implementations, compute budgets, VLM-judge choices, and dataset splits?
What assumptions explain the reported failure cascade from syntactically valid code to engineering-ready geometry, and which limitations are unmeasured?

CVAILGGRCLSEROCG

192 papers

system

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

40 Methodology quality 78 Recency 39 Reproducibility 27 Topical relevance

Literature review synthesis

Research Lines

Text/multimodal-to-CAD program generation

Translating natural language descriptions, images, or point clouds directly into executable parametric CAD programs, enabling rapid design prototyping and accessibility for non-experts.

Open edge: Robustness to ambiguous or underspecified prompts; guaranteeing that generated programs produce manufacturable and functionally correct assemblies; generalization to unseen design intents and complex multi-part structures.

Mesh/point cloud to parametric CAD reverse engineering

Recovering editable CAD construction sequences from 3D scans or meshes, which supports legacy part digitization and design reuse.

Open edge: Handling high geometric complexity, noise, and missing data; ensuring the reverse-engineered sequence is compact and semantically meaningful; validating on real-world scanned parts rather than synthetic meshes.

Unified benchmarks and evaluation protocols for CAD generation

Providing standardized multi-modal datasets, metrics, and evaluation pipelines that enable fair comparison across CAD generation and reconstruction methods.

Open edge: Whether geometric and validity metrics correlate with downstream engineering usability; the absence of multi-stage benchmarks that assess manufacturability, functionality, and assemblability at scale; the reliability of automated judges versus human expert evaluation.

Interactive and agentic CAD generation with clarification

Improving generation quality under ambiguous user prompts by having an agent proactively ask for clarification before code synthesis.

Open edge: Optimal trade-off between interaction cost and output quality; how to extend the clarifying dialogue to cover complex engineering constraints; and whether the approach scales to multi-turn, collaborative design workflows.

Shared Direction

Parametric CAD programs (construction sequences, B-Rep-based scripts) are the preferred output representation because they preserve editability and manufacturing history.
Geometric fidelity and executability are the primary evaluation dimensions; most works report intersection-over-union, Chamfer distance, and invalid ratio.
Large language models and vision-language models are the dominant backbone, often fine-tuned on curated CAD datasets.
There is a growing recognition that purely geometric metrics are insufficient and that benchmarks must incorporate design-intent alignment, manufacturability, and functional requirements.

Key Differences

Whether to generate CAD programs in a single pass or via iterative optimization and stepwise verification; one-shot methods are simpler but often less accurate, while stepwise methods improve fidelity at higher computational cost.
The choice of CAD scripting language and representation: some works target CadQuery, others use custom construction trees or feature-based operations, which affects portability and compatibility with existing CAD tools.
The granularity of evaluation: some benchmarks focus on single-part geometry, while others push toward multi-part assemblies with engineering constraints, leading to different conclusions about model capability.
The role of interaction: systems like ProCAD introduce explicit clarification loops, whereas most other works assume the input is self-contained, revealing a split on whether user-in-the-loop is necessary for practical deployment.

Open Questions

Can current text-to-CAD systems produce outputs that satisfy real manufacturing constraints such as tolerances, material properties, and assembly fit without human post-processing?
How reliable are VLM-based judges for assessing design intent, and can their judgments be calibrated against domain expert evaluations?
What is the minimal set of clarifying interactions needed to resolve most practical ambiguities in text-to-CAD, and does proactive clarification improve outcomes in a real user study?
To what extent do mesh-to-CAD reverse engineering methods generalize to scanned industrial parts with defects, occlusions, and non-ideal topology?
How can the community build a living benchmark that evolves with CAD toolchains and captures the diversity of mechanical, architectural, and electronic design domains?

CVAILGGRSECLHCSY

146 papers

application

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

44 Methodology quality 77 Recency 39 Reproducibility 28 Topical relevance

Literature review synthesis

Research Lines

Parametric mesh/image-to-CAD reverse engineering

Recovers editable parametric CAD construction sequences from meshes, images, or orthographic views by optimizing operations against geometric feedback or by stepwise decision-making.

Open edge: Generalization to real-world scans, sketches, and complex topologies; whether high geometric accuracy also preserves downstream editability and valid solid semantics.

Multimodal and code-generating CAD models

Maps text, image, sketch, or point cloud inputs to editable CAD code or unified multi-task CAD outputs in a single model.

Open edge: Real-world robustness beyond synthetic image-code pairs, unseen operations, and domain constraints such as manufacturing tolerances or CAE/CAM requirements.

Long-sequence generative CAD program modeling

Models parametric CAD sequences with latent or state-space generative architectures to produce longer valid command sequences.

Open edge: Whether longer sequence generation corresponds to meaningful design complexity, parametric validity, and downstream task utility; dataset coverage and operation semantics remain lightly verified.

Evolutionary and synthetic CAD data generation

Creates complex CadQuery programs through VLM-guided iterative edits and validation, then uses augmented script-geometry pairs to train downstream image-to-CAD models.

Open edge: Generated data may overfit to programmatic realism; reproducibility, diversity, and industrial-grade constraint coverage are not yet established.

Shared Direction

Parametric CAD should be represented as editable programs or code, not only as meshes or solids, to preserve design intent.
Geometric feedback, whether through IoU, Chamfer distance, or executable assertions, is central for training and evaluation.
Recent methods focus on CadQuery or similar code-style outputs and aim to improve validity and geometric similarity over task-specific baselines.
Current evidence relies heavily on synthetic benchmark datasets such as DeepCAD, Fusion 360, MCB, or newly created CAD program datasets.
There is a shift toward multimodal and unified models that handle several CAD tasks in a single framework.

Key Differences

Representation differs: some generate structured CAD construction sequences, others generate Python-style CAD code, and others use latent sequence models.
Supervision differs: hybrid optimization uses geometric objectives, stepwise methods use intermediate geometry comparison or RL rewards, and others use VLM-guided edits or program evolution.
Evaluation target differs: syntax validity, 3D solid similarity, invalid ratio, benchmark-specific SOTA, or design-intent adherence via geometric assertions.
Interaction mode differs: some methods reconstruct from static mesh/image input, while others target conditional program completion with explicit design headers.
Deployment assumptions differ: universal end-to-end models assume task unification benefits, whereas stepwise or hybrid approaches assume interpretable intermediate geometry improves reconstruction.

Open Questions

Do gains on synthetic CAD reconstruction benchmarks translate to improved mechanical design, manufacturing, CAE/CAM, or 3D printing workflows?
Can current mesh/image-to-CAD methods handle noisy real-world scans, hand sketches, unseen operation sequences, and partially specified designs?
How can evaluation avoid circularity when design-intent header extraction and metric scoring are based on the same generated code or geometry?
Which representation, code-based or parametric sequence-based, better preserves editability, parametric constraints, and downstream interoperability?
Does longer or more complex program generation reflect meaningful design complexity, or does it exploit dataset-specific sequence patterns?
What verification evidence is missing for reproducibility, given that code, datasets, baselines, and limitation reports are not consistently available in the packet?

CVLGAIGRSECLCGHC

123 papers

theory

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

53 Methodology quality 81 Recency 37 Reproducibility 33 Topical relevance

Literature review synthesis

Research Lines

Benchmark-driven multimodal CAD program evaluation (e.g., CADBench, MUSE, BenchCAD)

Systematic comparison of programmatic CAD generation across input modalities, metrics, and design criteria beyond geometric similarity.

Open edge: Whether benchmark gains transfer to unseen part families and align with formal design reasoning remains unclear.

LLM-driven text-to-CAD program synthesis with grounding

Generating executable parametric CAD scripts from language and connecting references to B-rep primitives or parameters.

Open edge: Ambiguity handling, fine-grained parameter accuracy, and generalization beyond curated distributions remain partially open.

Self-supervised and sequence-based parametric CAD generative models

Learning long parametric CAD sequences or editable B-rep surfaces from point-level or reconstruction supervision.

Open edge: Real-world boundary-representation complexity and cross-domain transfer beyond ABC/Objaverse-style data need more evidence.

Engineering-intent evaluation and agentic clarification

Assessing manufacturability, functionality, and assemblability while reducing prompt ambiguity before CAD code synthesis.

Open edge: Scalable reliable judges and low-overhead clarification across broader design tasks are not yet established.

Shared Direction

Parametric CAD generation is converging on executable program or B-rep representations to preserve editability and manufacturability.
Multiple papers agree geometric fidelity alone is insufficient; executability, design intent, manufacturability, and assemblability must also be evaluated.
Strong current models still fail on fine-grained structure, complex geometry, or unseen part families despite high benchmark-oriented performance.
Grounding language or visual inputs to geometric primitives, structure trajectories, or parameters is treated as a key mechanism for consistency and accuracy.

Key Differences

Representation choice: CadQuery script generation, latent state-space sequences, and differentiable Bezier/B-rep patches lead to different editability and supervision trade-offs.
Supervision signal: point-level self-supervision on unannotated 3D data differs from curated text or specification supervision with or without reinforcement learning.
Evaluation target: one line emphasizes geometric fidelity and user preference, another emphasizes manufacturability and assemblability, and another emphasizes robustness to ambiguous prompts.
Interaction mode: some systems insert a proactive clarification agent before synthesis; others generate directly from a single prompt.
Training mechanism: GAN-based latent generation, supervised fine-tuning plus reinforcement learning, and ranking-based structure or parameter grounding reflect different assumptions about learning CAD structure.

Open Questions

How well do current empirical CAD generation results generalize to unseen industrial part families and to shifts across clean meshes, noisy meshes, single-view, photorealistic, and multi-view inputs?
Does the formal geometric modeling theory implied by the topic align with observed failure cascades from executable code to valid geometry to engineering-ready design?
Can proactive clarification, hierarchical reasoning, or parameter grounding scale while keeping interaction overhead low and avoiding shortcut learning?
What evidence is missing because several top-paper analyses omit limitation details or reproducible code and dataset access despite methodology and reproducibility scores?
Can benchmark ranking instability across metrics be resolved into a stable definition of CAD generation quality?

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