Aggregate analysis

Research Analysis

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

280 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

LLM-driven parametric CAD generation from text or multimodal inputs

Translates natural language or multimodal prompts into executable parametric CAD programs (e.g., CadQuery scripts, construction trees) to automate design synthesis.

Open edge: Generated programs often fail to satisfy engineering constraints like manufacturability or assemblability, and robustness to ambiguous prompts remains limited.

Optimization-based mesh-to-CAD reconstruction

Recovers editable parametric CAD sequences from discrete 3D data (meshes, point clouds) by incrementally fitting operations and validating via geometric feedback.

Open edge: Scalability to complex topologies and the balance between reconstruction accuracy and program validity are unresolved; generalization across object categories is underexplored.

Unified multi-modal multi-task benchmarks and evaluation protocols

Provides standardized evaluation across datasets, modalities, and tasks to surface failure modes and compare model capabilities beyond geometric similarity.

Open edge: Current metrics insufficiently capture engineering readiness (e.g., functional play, assemblability); automated judges only partially align with human expert evaluation.

Interactive and proactive clarification agents for CAD

Improves robustness by allowing models to query users for specification clarification before generating CAD programs, reducing ambiguity-driven errors.

Open edge: The trade-off between interaction cost and generation quality, and the generalization of clarification strategies to diverse design contexts, are not well understood.

Shared Direction

Parametric CAD representations (B-Rep or construction sequences) are essential for editability and downstream engineering use.
Program executability and volumetric similarity (e.g., IoU, Chamfer distance) serve as primary evaluation metrics, though they are recognized as insufficient for practical design.
Pure end-to-end generation struggles with complex geometries and engineering constraints; hybrid approaches (optimization + learning, or agentic workflows) are seen as promising.
Multi-modal inputs (text, images, mesh, point clouds) are increasingly supported, reflecting real-world design workflows.

Key Differences

Representation: Some methods use CadQuery or domain-specific scripting languages, while others adopt hierarchical construction trees or unstructured text-based CAD primitives.
Supervision: Training strategies differ—supervised fine-tuning on curated datasets, reinforcement learning for geometric alignment, and hybrid optimization with geometric feedback.
Evaluation target: Some benchmarks emphasize geometric fidelity and program validity, while others target engineering-grade criteria like manufacturability, functionality, and assemblability.
Interaction mode: Most systems generate in one pass, but ProCAD introduces a proactive clarification loop, shifting the assumption from fully specified to underspecified inputs.

Open Questions

Can generated CAD models be automatically validated against engineering constraints (e.g., stress, manufacturing cost) without expert human review?
How can models generalize from limited parametric CAD data to novel object categories or highly complex assemblies?
What is the optimal balance between automated generation and human-in-the-loop clarification for design tasks?
To what extent do current metrics predict real-world usability and downstream task success (e.g., simulation, fabrication)?

CVAILGGRCLSEROCG

230 papers

benchmark

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

44 Methodology quality 79 Recency 38 Reproducibility 28 Topical relevance

Literature review synthesis

Research Lines

Benchmark task and dataset design

Creates shared, standardized evaluation settings for AI-assisted CAD generation, reverse engineering, and question answering, enabling fair comparison across methods.

Open edge: Whether the task distributions and object categories cover the full range of real-world mechanical design, including complex assemblies, noisy inputs, and engineering constraints such as tolerances and material properties.

Metric and stress-test design

Probes prediction quality, controllability, and robustness beyond aggregate scores by stratifying samples by geometric complexity, measuring multiple failure modes, and evaluating design intent alignment.

Open edge: The alignment between these proxy metrics and the actual downstream decisions engineers care about (e.g., manufacturability, cost) has not been empirically validated, and the sensitivity of VLM judges to domain shift is unclear.

Comparable baselines and reporting

Provides unified datasets, evaluation protocols, and public leaderboards that make results auditable and models comparable under consistent conditions.

Open edge: Stability of benchmark conclusions across different implementations, compute budgets, and the reproducibility of LLM-based judges has not been systematically studied.

Shared Direction

Evaluation is moving from single geometric metrics (e.g., Chamfer distance) to multi-dimensional assessments that include executability, program compactness, and engineering criteria.
Staged evaluation protocols are adopted to isolate failures in code syntax, shape validity, and design intent, rather than collapsing everything into one score.
Automated judges based on large language or vision-language models are increasingly used to assess semantic and functional aspects of CAD models at scale.
Benchmarks are designed to expose specific failure modes—such as brittleness under modality shift and degradation with geometric complexity—rather than just ranking models.

Key Differences

Task definitions diverge: some benchmarks focus on generative program synthesis from text or images, while others target reverse engineering of meshes into parametric sequences, leading to incompatible evaluation setups.
Input modalities vary from single-view images and point clouds to multi-view renders and sketches, making cross-benchmark comparisons difficult and raising questions about which modality best reflects real-world use.
The definition of a successful CAD model differs: one line of work treats geometric fidelity as the primary goal, while another emphasizes design intent, manufacturability, and assemblability, often requiring human or VLM judgment.
Interaction modes are treated differently: most benchmarks assume a single-turn generation, but one line of work introduces proactive clarification agents, challenging the assumption that prompts are self-contained specifications.

Open Questions

How reliably can VLM-based judges assess engineering criteria such as functionality and manufacturability, and how do they compare to expert human evaluation?
Do current benchmarks cover the diversity of real manufacturing constraints, including tolerances, material choices, and assembly processes, or are they limited to idealized geometric shapes?
Can findings from one benchmark (e.g., mesh-to-CAD) transfer to another (e.g., text-to-CAD) when the task definitions and success metrics are fundamentally different?
How stable are the reported model rankings under changes in implementation details, compute budgets, or the specific subset of samples used for evaluation?

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

143 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 CAD reverse engineering

Recovering editable CAD construction sequences or programs from meshes, point clouds, or images, enabling redesign and manufacturing reuse.

Open edge: Scalability to complex industrial parts, handling of unseen operations and noisy real-world scans, and validation that the reconstructed program reflects true design intent rather than merely approximating geometry.

End-to-end multimodal CAD generation

Generating CAD code or construction sequences directly from text, images, or sketches using large vision-language models, unifying diverse CAD tasks in a single framework.

Open edge: Robustness to out-of-distribution inputs, syntax correctness and semantic validity of generated code beyond training categories, and the ability to guarantee manufacturability and adherence to implicit design rules.

Controllable generation and design intent

Diagnosing how conditioning signals like design-intent headers influence CAD program generation, and developing methods to ensure faithful adherence to specified constraints.

Open edge: Causal identification of true header effects versus distributional artifacts, generalization across different base models and CAD domains, and the design of evaluation metrics that are independent of the conditioning mechanism.

Shared Direction

Editable parametric CAD is the target representation, not static meshes or point clouds, because it supports downstream engineering modifications.
Multi-modal inputs (images, meshes, point clouds, text) are increasingly used to drive CAD generation, moving beyond single-modality reconstruction.
Evaluation must go beyond geometric similarity (IoU, Chamfer distance) and include executable validity, syntax correctness, and design-intent adherence to be meaningful for real workflows.

Key Differences

Representation choices diverge: some works reconstruct CAD construction sequences as step-by-step operations, while others generate executable scripts in a general-purpose CAD language (CadQuery).
Supervision strategies differ: stepwise geometric feedback at each construction step versus end-to-end training from image–code pairs without intermediate supervision.
Architectural preferences vary: hybrid optimization, decision transformers with offline RL, state-space models (Mamba), and fine-tuned vision-language models all compete, with no clear winner for long-sequence or complex designs.
Evaluation scope differs: some emphasize geometric accuracy on standard benchmarks, while others highlight syntax validity, dataset creation, or causal analysis of conditioning signals, making cross-comparison difficult.

Open Questions

How well do current methods generalize to real-world photographs, hand sketches, or incomplete scans, given experiments on curated datasets?
Can the evaluation metrics be made independent of the header or condition extraction process to avoid circularity, as suggested by the derangement-control study?
What is the minimal set of design constraints (dimensions, tolerances, material) that must be satisfied for generated CAD to be considered manufacturing-ready, and how should they be verified?

CVLGAIGRSECLCGHC

119 papers

theory

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

53 Methodology quality 80 Recency 37 Reproducibility 33 Topical relevance

Literature review synthesis

Research Lines

Multimodal procedural CAD generation

Generates editable B-Rep models from text, image, or point cloud inputs by emitting executable CAD programs, enabling direct editing and high geometric fidelity without CAD-specific annotations.

Open edge: Models often fail on complex features like sweeps and lofts, exhibit brittleness under modality shift, and show limited generalization to distribution-shifted part families.

Manufacturability-aware CAD benchmarks

Provides standardized evaluation protocols that go beyond geometric similarity to assess executability, functionality, manufacturability, and assemblability, revealing failure cascades and ranking shifts across metrics.

Open edge: Automated VLM-based judges need further validation against human expert assessments; benchmarks currently focus on single parts, leaving multi-part assembly evaluation unaddressed.

Interactive clarification and design intent alignment

Improves robustness to ambiguous prompts by proactively asking clarification questions or aligning hierarchical structure, reducing invalidity rates and parameter errors.

Open edge: The impact of interaction overhead on user efficiency and the scalability to multi-turn, multi-component designs remain unexplored.

Shared Direction

B-Rep is the preferred editable representation for CAD generation.
Program synthesis (CadQuery scripts or parametric sequences) enables precise and editable model creation.
Multimodal input (text, image, point cloud) is essential for flexible CAD authoring.
Evaluation must include multiple dimensions: geometric fidelity, executability, and engineering constraints such as manufacturability.

Key Differences

Program representation differs: some use CadQuery scripts, others use custom parametric sequences (e.g., DeepCAD-style).
Training strategies vary: supervised fine-tuning is common, but reinforcement learning is used in some cases to improve generalization.
Approaches to complex geometry differ: some methods handle only sketch-and-extrude, while others attempt sweeps and lofts with mixed success.
Evaluation frameworks choose between design-specific rubrics with VLM judges and broader metrics, introducing variation in what constitutes a good model.

Open Questions

How can generative models reliably produce high-level operations like sweeps and lofts while maintaining executability?
What methods can guarantee that generated models meet real-world manufacturing constraints (e.g., tolerances, material properties)?
How can we align automated evaluation metrics with expert engineering judgment, especially for design intent?
Can the clarifying or alignment strategies scale to multi-part assemblies and interactive design workflows?

CVAILGGRCLSECEMA

50 papers

survey

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

13 Methodology quality 76 Recency 51 Reproducibility 10 Topical relevance

Literature review synthesis

Research Lines

Taxonomy and roadmap synthesis

Organizes fragmented AI4CAD methods into coherent directions such as generation, reconstruction, and interaction, revealing research gaps and milestones.

Open edge: Whether the taxonomy is backed by systematic evidence across all subfields or relies on a selective sample of recent papers; missing coverage of traditional CAD paradigms.

Comparative evaluation synthesis

Summarizes how different evaluation choices (e.g., geometric fidelity vs. topological validity) shape conclusions, as seen in BrepForge's comparison against baselines.

Open edge: Results are often not comparable across representations or datasets; no unified benchmark exists for cross-task evaluation, and metric sensitivity to hyperparameters is underexplored.

Open problem and gap mapping

Identifies concrete next questions such as controllable generation, real-world sim-to-real transfer, and long-horizon CAD interaction learning from video.

Open edge: Prioritization of which gaps (e.g., topological guarantees vs. dataset scale) matter most for practical CAD adoption remains unclear.

Shared Direction

Deep learning is the primary tool for CAD-related tasks, replacing hand-crafted geometric algorithms.
Factorized or multi-stage pipelines are preferred to handle the complexity of CAD structures.
Synthetic data plays a crucial role in training due to the scarcity of annotated real CAD data.
Code and dataset release is becoming common practice, improving reproducibility.

Key Differences

Choice of 3D representation: B-rep offers engineering precision but is harder to generate, while meshes and point clouds are easier but lack topological guarantees.
Supervision paradigm: some methods use reinforcement learning for alignment, others use autoregressive or GAN-based generation, leading to different stability and quality trade-offs.
Evaluation focus: geometric accuracy vs. topological validity vs. interaction success rate, with no consensus on a primary metric.
Interaction modality: some works target automatic generation, others learn from human UI interactions, reflecting different deployment assumptions.

Open Questions

How can we design a unified evaluation protocol that fairly compares B-rep, mesh, and point cloud generation methods?
What is the minimal real-world supervision needed to bridge the sim-to-real gap in CAD model generation and reconstruction?
Can factorized generation approaches guarantee global topological consistency without sacrificing geometric detail?
How should long-horizon CAD interaction models be evaluated beyond imitation accuracy, e.g., on task completion or design intent preservation?

GRROCVAI