Paper deep dive
Vision2CAD: A Visual Agent Harness for Explicit Geometry Referencing and Localization in Parametric CAD Modeling
Xi Cheng, Chenxi Zhai, Hang Cheng, Mingyu Fan, Pingfa Feng, Long Zeng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/22/2026, 5:04:45 AM
Summary
The paper introduces Vision2CAD, a visual agent harness for generating parametric CAD models from images by combining Vision-Language Model (VLM) reasoning with deterministic CAD kernel operations. It addresses key challenges in geometry referencing, sketch-plane coordinate understanding, and external sketch constraints through an ID-based interface, a local-coordinate bridge, and projected-edge localization. The authors also present the Geometry Explicit Reference Dataset (GERD) and demonstrate improved performance metrics (mIoU and Chamfer distance) over existing methods like DeepCAD.
Entities (7)
Relation Signals (6)
Vision2CAD → improves → mIoU
confidence 95% · Vision2CAD improves mIoU by 11.1% and 5.6%...
Vision2CAD → reduces → Chamfer Distance
confidence 95% · reduces Chamfer distance by 17.3% and 41.8%, respectively.
Vision2CAD → uses → VLM
confidence 95% · Vision2CAD... combines vision-language model (VLM) reasoning with deterministic CAD kernel operations.
Vision2CAD → outperforms → DeepCAD
confidence 92% · On GERD-EVL and a DeepCAD test subset, Vision2CAD improves mIoU by 11.1% and 5.6%...
Vision2CAD → addresses → geometry referencing
confidence 90% · To address these limitations, we focus on three problems... explicit geometry referencing...
Vision2CAD → utilizes → GERD
confidence 90% · We also introduce the Geometry Explicit Reference Dataset (GERD)... On GERD-EVL... Vision2CAD improves...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generating parametric CAD models requires accurate geometry and stable feature dependencies. Existing methods face challenges in selecting geometric references, interpreting sketch-plane local coordinates, and establishing sketch constraints to projected external geometry. We present Vision2CAD, a visual agent harness that combines vision-language model (VLM) reasoning with deterministic CAD kernel operations. An ID-based interface supports explicit geometry selection, a local-coordinate bridge converts view coordinates into sketch coordinates, and projected-edge localization supports external sketch constraints. These mechanisms establish feature dependencies within the supported modeling operations and constraint types. We also introduce the Geometry Explicit Reference Dataset (GERD), which aligned commands, geometry states and IDs at every modeling step. On GERD-EVL and a DeepCAD test subset, Vision2CAD improves mIoU by 11.1\% and 5.6\% and reduces Chamfer distance by 17.3\% and 41.8\%, respectively. Parameter-editing experiments and ablation studies further proved the preservation of parametric dependencies.
Tags
Links
- Source: https://arxiv.org/abs/2609.22688v1
- Canonical: https://arxiv.org/abs/2609.22688v1
Trouble viewing inline? Open PDF directly →
Full Text
48,547 characters extracted from source content.
Expand or collapse full text
Vision2CAD: A Visual Agent Harness for Explicit Geometry Referencing and Localization in Parametric CAD Modeling Xi Cheng a , Chenxi Zhai a , Hang Cheng a , Mingyu Fan a , Pingfa Feng a,b , Long Zeng a,∗ a Division of Advanced Manufacturing, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, 518055, China b Department of Mechanical Engineering, Tsinghua University, Beijing, 100084, China Abstract Generating parametric CAD models requires accurate geometry and stable feature dependencies. Existing methods face challenges in selecting geomet- ric references, interpreting sketch-plane local coordinates, and establishing sketch constraints to projected external geometry. We present Vision2CAD, a visual agent harness that combines vision-language model (VLM) reasoning with deterministic CAD kernel operations. An ID-based interface supports explicit geometry selection, a local-coordinate bridge converts view coordi- nates into sketch coordinates, and projected-edge localization supports ex- ternal sketch constraints. These mechanisms establish feature dependencies within the supported modeling operations and constraint types. We also introduce the Geometry Explicit Reference Dataset (GERD), which aligned commands, geometry states and IDs at every modeling step. On GERD-EVL and a DeepCAD test subset, Vision2CAD improves mIoU by 11.1% and 5.6% and reduces Chamfer distance by 17.3% and 41.8%, respectively. Parameter- editing experiments and ablation studies further proved the preservation of parametric dependencies. ∗ Corresponding author. Email addresses:chengxi23@tsinghua.org.cn(Xi Cheng), dcx24@mails.tsinghua.edu.cn(Chenxi Zhai),chenghan24@mails.tsinghua.edu.cn (Hang Cheng),my-fan25@mails.tsinghua.edu.cn(Mingyu Fan), fengpf@tsinghua.edu.cn(Pingfa Feng),zenglong@sz.tsinghua.edu.cn(Long Zeng) arXiv:2609.22688v1 [cs.CV] 19 Sep 2026 Keywords:Engineering informatics, Computer-aided design, Artificial intelligence, Knowledge-intensive engineering 1. Introduction Parametric CAD models represent products through an ordered modeling feature sequence, whose dimensions, constraints, and geometric references en- code design intends. When an upstream parameter changes, parameter and constraint dependencies allow downstream features to update while preserv- ing the design intends [ 1,2]. This editability makes parametric CAD mod- els fundamental to iterative design, engineering analysis, and manufactur- ing. Meanwhile, product images, renderings, and design sketches are widely available during engineering design and product development. Automatically converting such visual inputs into parametric CAD models can reduce the manual effort required for CAD reconstruction. The resulting models can then be directly edited and reused in downstream engineering tasks, includ- ing analysis, optimization, and manufacturing [ 3,4]. Early Transformer- or diffusion-based methods typically represent para- metric CAD as predefined command sequences. Representative methods such as DeepCAD generate sketches and extrusion parameters, but their represen- tations generally lack references to existing faces, edges, or vertices, making it difficult to express dependencies between features [ 3,5–7]. Subsequent LLM- and VLM-based methods generate executable CAD programs, such as CadQuery scripts. These programs support a wider range of operations and allow limited geometry referencing. However, methods that generate a com- plete program without intermediate execution cannot observe the evolving geometry states, making reference selection difficult to verify and correct [ 8– 10]. Recent agent-based methods use execution results and visual feedback to iteratively verify and revise generated models [ 11–13]. These methods improve geometric reconstruction and enable correction of some referencing errors, while related state-aware methods further support explicit entity se- lection [14,15]. Nevertheless, reliable geometry referencing across modeling operations remains challenging, and its integration with sketch-plane local coordinates understanding and sketch constraints to projected external ge- ometry remains incomplete. As a result, although the generated CAD mod- els may reproduce the target appearance, it fails to establish the feature dependencies required to preserve design intent (top of Figure 1). 2 Command Sequence-Based Method Vision2CAD AAAB ACADAE constraints: 1. tangent(A,AB) 2. tangent(A,AE) 3. ... Extrude geom_table: faceAK: boundary: [AG,AH,AI,AJ] edgeAG: samples_proj: [...] Extrude 1. Hard to bind the sketch plane to an arbitrary existing face. (Determine푂,+푍,+푋generally, not bound to existing face) Sketch Sketch Sketching on face AK AM AN AM.start AN.start AO AP constraints: 1. center(AO)=proj(AM.start) 2. center(AP)=proj(AN.start) 3. ... 2. Sketch-plane local coordinate system is not explicitly exposed. 3. Hard to establish constraints between sketch primitives and projected geometry. Extrude Extrude AG AH AI AJ Parameter Change Parameter Change Update violates the design intent. Update preserves the design intent. geom_table: edgeAM: samples_proj: [푥 1 ,푦 1 ,푥 2 ,푦 2 ,...,푥 푘 ,푦 푘 ] Figure 1:The advantages of Vision2CAD.Previous works often struggle to select arbitrary faces, edges, and vertices. Vision2CAD exposes entity IDs through ID-tagged views and a geometry table, allowing the VLM to directly select existing geometry. Sec- ond, the origin and axes of sketch-plane local coordinate is usually implicit, these methods cannot reliably generate the key-point coordinates required to define sketch primitives. Vision2CAD exposes view coordinates through ticks around the viewport, the CAD kernel transforms view coordinates to the sketch coordinates. Third, existing methods generally do not provide the sketch plane-projected locations of external geometry, making it dif- ficult to establish external sketch constraints. Vision2CAD provides projected samples and source IDs of existing edges, allowing generated sketch primitives align to projected geometry, constraints can then be searched and established. When upstream parameters change, above mechanism allows downstream features updates while maintaining the de- sign intend. To address these limitations, we focus on three problems in AI-based parametric CAD modeling: explicit geometry referencing, sketch-plane local coordinate understanding, and sketch constraints to projected external ge- ometry. Vision2CAD provides an ID-based geometry referencing interface, a local-coordinate bridge between view and sketch coordinates, and projected- geometry-aware constraint generation. These mechanisms are integrated into a visual agent harness that combines VLM reasoning with deterministic CAD kernel operations. The system generates native parametric CAD models with editable sketches, feature histories, and explicit dependencies within its sup- ported operation and constraint set (bottom of Figure 1). We also introduce the Geometry Explicit Reference Dataset (GERD) to support modeling with explicit geometry references. The main contributions of this work are as follows: •We identify three core challenges in AI-based parametric CAD gener- ation: geometry referencing, understanding sketch-plane local coordi- nate systems, and establishing sketch constraints to projected external 3 geometry. •We introduce Vision2CAD to address these challenges within a defined set of modeling operations and constraints. An ID-based visual inter- face and geometry table supports geometry selection, a local-coordinate bridge converts view coordinates into sketch coordinates, and projected- edge localization supports external sketch constraints. •We construct the GERD dataset, comprising 2,700 parts across seven operation categories. It aligns commands, geometric states, and explicit Body, Face, Edge, and Vertex IDs at every modeling step. •We evaluate static reconstruction and a uniform 10% increase in extru- sion depths on GERD-EVL and a DeepCAD test subset. Vision2CAD achieves the best performance among the evaluated methods. Ablation studies support the contributions of the proposed mechanisms. 2. Related Work 2.1. Structured Sequence Generation with Specialized Algorithm Early methods use Transformers or diffusion models to generate CAD models as sequences of predefined commands and numerical parameters. DeepCAD represents sketch primitives, sketch planes, and extrusion pa- rameters as command vectors and learns their sequential structure with a Transformer [ 5]. Text2CAD introduces text-conditioned sequence generation, while CAD-Diffuser reconstructs construction sequences from point clouds through multimodal diffusion [ 6,7]. Image-conditioned methods, including Img2CAD and CADCrafter, further connect visual inputs to structured CAD generation [3,4]. These methods provide compact representations for learning sketch-and- extrude construction histories. However, their command vocabularies pri- marily describe what operation to perform and its numerical parameters, without explicitly identifying existing faces, edges, or vertices as references. Consequently, reproducing the target shape does not ensure that downstream features follow upstream changes as intended. 2.2. CAD Code Generation with LLMs and VLMs LLM- and VLM-based methods generate executable modeling code, allow- ing them to use the operations and referencing mechanisms of CAD libraries 4 such as CadQuery. CAD-Recode translates point clouds into Python CAD programs [8], and Ortho2CAD generates CadQuery code from orthographic drawings [9]. CAD-Coder learns image-to-code generation, while ReCAD incorporates reinforcement learning to improve parametric CAD generation with VLMs [10,16]. Compared with predefined sketch-and-extrude sequences, CAD programs can express a broader range of operations and access some existing geometry through selectors [17]. However, most existing methods generate the com- plete program without observing intermediate CAD states. This requires the model to implicitly infer the evolving geometry, making entity selection error-prone. 2.3. Parametric CAD Generation with Agents Agent-based methods incorporate execution, inspection, and revision. CAD-Assistant provides FreeCAD tools and visual feedback, while Seek- CAD combines stepwise feedback with references to start, end, and swept faces [ 11,12]. Multi-agent systems include CADSmith, which combines error correction, geometric validation, and VLM assessment, and Fusion AI-AD, which integrates design knowledge processing with SolidWorks modeling [ 13,18]. These works improve reconstruction and allow some referencing errors to be corrected. Nevertheless, successful execution and appearance matching do not verify intended feature dependencies. Other problems such as sketch-plane local coordinate understanding and sketch plane projected-geometry dependencies are not jointly resolved. 3. Core Challenges in AI-Based Parametric CAD Modeling 3.1. Geometry Reference Geometry reference involves selecting specific existing faces, edges, or vertices for a modeling operation. For example, creating a sketch requires selecting its support plane. Incorrect references may produce the intended initial shape but lead to unintended updates after parameter edits ( Figure 2). Early methods predict sketch-plane origins and axis directions di- rectly [ 5]. Later approaches add limited dependencies, such as referencing a face through the sketch edge that generates it [12,17,19]. Richer op- erations require more general geometry references. Engineers select these 5 (A1)(A2) (a1)(a2) (B1)(B2) (b1)(b2) Figure 2:Geometry reference.With correct references, the sketch and swept solid in (A1) and (B1) update correctly in (A2) and (B2). Without references, the initially identical geometry in (a1) and (b1) fails to update correctly in (a2) and (b2). entities interactively; an AI system needs a representation that makes them identifiable to the VLM. 3.2. Understanding Sketch-Plane Local Coordinate Systems Sketch primitives are defined in the sketch plane’s two-dimensional local coordinate system. The CAD kernel assigns the origin and axes. Identical coordinates on different planes can therefore produce very different spatial positions ( Figure 3). 0.10.30.5 0.1 0.2 0.4 0.10.30.5 0.1 0.2 0.4 (a1)(a2) (b1)(b2) Figure 3:Sketch-plane local coordinates.Cyan marks the sketch planes; black trian- gles are the sketches. The front views in (a2) and (b2) show identical local definitions, but the spatial positions in (a1) and (b1) differ. Many existing methods primarily use sketch planes parallel to the XOY, YOZ, or XOZ planes, with local axes aligned or oppositely with the global 6 axes [5,19,20]. SkexGen reports that such planes account for over 99% of DeepCAD sketches [21], reducing the need to handle more varied sketch- plane orientations in that dataset. Nevertheless, inclined planes occur in engineering models. The challenge is to expose their coordinates through a clear visual interface or map predicted positions into sketch local coordinate system. 3.3. Sketch Constraints with Projected External Geometry Projected geometry constants link sketch primitives to existing edges or vertices. For example, a circle center can coincide with a projected vertex. This association lets the sketch update when upstream dimensions change ( Figure 4). existing edge projected edge 푒 푝푟표푗 sketch primitive line 푒 푠푘ℎ sketch primitive circle 푐 푠푘ℎ existing vertex projected vertex 푣 푝푟표푗 constraints: 1. coincide(푣 푝푟표푗 , center(푐 푠푘ℎ )) 2.coline(푒 푝푟표푗 , 푒 푠푘ℎ ) (a)(b)(c) Figure 4:Sketch constraints with projected geometry.(a) Sketch primitives are constrained to projected geometry. (b, c) They update with that geometry after upstream parameter changes. Several methods output sketch primitive definitions directly [5,12,19,21]. Pointer-CAD supports snapping to existing edges [14]; maintaining such an association during later edits additionally requires a persistent constraint. Establishing constraints to projected external geometry requires identifying both the source geometry and its projection on the sketch plane. 4. Method Vision2CAD constructs a parametric CAD model from an input image by combining VLM predictions with deterministic CAD kernel operations. Its interfaces support entity referencing, sketch coordinate conversion, and external sketch constraints within the supported modeling operations. The overall workflow is shown in Figure 5. 7 stage1: selecting (determine three components)stage2: sketchingstage3: verify document 퐷 푡 type reference parameter sketch planesweep pathrevolve axisloft section extrude sweep revolve loft extrude depthrevolve anglefillet radius arc circle line document 퐷 푡+1 FreeCAD 퐼 ∗ 퐷 푡 퐼 푡 depth 퐼 푡 ret RAG Register 푆 푡 VLM 퐼 푡 tag 푋 푡 1 푇 푡 geom 퐼 푡 tag,s 푇 푡 geom,s VLM 푋 푡 2 퐷 푡 퐷 푡+1 푋 푡 1 VLM 퐼 푡 pv 푋 푡 3 FreeCAD 퐼 ∗ 퐷 0 퐷 1 퐷 2 퐷 3 퐷 4 퐷 5 퐷 6 Figure 5: Vision2CAD workflow. Given a target image I ∗ , Vision2CAD performs Nmodeling steps and outputs a FreeCAD documentD N . Each step comprises three stages.Stage 1:The depth imageI depth t of current documentD t is registered toI ∗ , the result view provides tag imageI tag t , geometry tableT geom t , and a pseudo-bounding sphere S t .I depth t andI ∗ are used to retrieve modeling examplesI ret t .I tag t andT geom t provide geometry IDs for explicit geometry reference. The VLM outputX (1) t contains all data required for modeling command except sketch.Stage 2:S t helps to determine suitable sketch viewport,I tag,s t andT geom,s t are generated under this viewport.I tag,s t andT geom,s t provide projected external edge localization, which guides primitive placement, the CAD kernel converts view coordinates into sketch coordinates and adds supported constraints. Stage 3:FreeCAD executes the modeling command from Stage 1,2 and renders modeling previewI pv t . The VLM assesses whether it adds a missing target structure or a useful intermediate feature. Accepted commands updateD t toD t+1 ; rejected commands return to Stage 1 with rejection reasons. VLM context is retained within each modeling step and reset between steps. 8 4.1. Stage 1: Operation Selection 4.1.1. Registration We render a depth imageI depth t fromD t . Relative linear depth is shown from white (near) to black (far). Magenta solid and dashed lines mark visible and occluded edges. The default PNP view has direction (−0.577,0.577,−0.577)and up vector(−0.408,0.408,0.816). Registration adjusts the camera scale and translation to alignI depth t with I ∗ . The resulting cameraC reg t , places existing geometry at the corresponding target location, helping identify regions to be modeled in Stage 1 (Figure 6 left). 퐼 푡 depth 퐼 ∗ registrationadjust 퐷 푡 camera by registration result registered camera 퐶 푡 reg make pseudo- bounding sphere 푆 푡 sketching to build target feature in green frame using 푆 푡 to obtain suitable sketch viewport using 퐷 푡 to obtain suitable sketch viewport Figure 6:Registration and sketch-view framing.In Stage 1 (left), registration aligns the current geometry with the target. In Stage 2 (right), a pseudo-bounding sphereS t , constructed fromD t underC reg t , helps keep the intended sketch within the viewport. Fitting the view to the existing solid alone can exclude part of the new sketch, as illustrated by the feature in the green frame. IncludingS t inFitAll()provides additional space. The sphere’s center lies on a symmetry plane of the solid’s bounding box, and its bounding box contains that of the solid. The sphere is chosen to limit changes to the registered camera’s scale and translation whenFitAll()is applied. UsingC reg t , we render the tag imageI tag t and export the geometry table T geom t . They expose the current geometry together for reference selection ( Figure 7). 4.1.2. Retrieval-Augmented Generation (RAG) We use the current depth imageI depth t and target imageI ∗ to retrieve modeling examplesI ret t . The retrieved pseudo-previews provide visual prece- dents for the next operation (Figure 8). 9 퐼 푡 tag Edge ID Tag Tag leading line Leading line anchor View coordinate grid tick 푇 푡 geom "Edge": "AC": "samples": [[0.18, 0.65], ... ], "type": "line" , "AM": ..., ... , "Face": "A": "edges": ["AC", "AM", "AG", "AH"], "type": "plane" , "AB": ..., ... Sample points’ coordinate in view coordinate system Edge type Edge ID Face ID Face boundary edges’ ID Face type Figure 7:Tag imageI tag t and geometry tableT geom t .I tag t shows depth, coordinate ticks, and edge IDs. Visible and occluded edges use solid and dashed magenta lines; vertices use magenta squares and asterisks. Labels omit some short edges and edges invisible in hidden-line removal mode.T geom t lists all edge types and uniformly sampled view coordinates. Face entries contain surface types and boundary edge IDs. The VLM selects faces and edges by ID, and vertices byID Edge .StartorID Edge .End. Edge samples are ordered from start to end. Tag leader lines are anchored at random positions within [0.1,0.25]of the normalized edge parameter domain, typically closer to the start. These cues distinguish the endpoints. Similarity to retrieval recordiis a weighted combination of cosine simi- larities: s depth i =φ ( I depth t ) T φ ( I depth i ) , s target i =φ(I ∗ ) T φ(I ∗ i ), s i =λs depth i + (1−λ)s target i ,0≤λ≤1. Here,s depth i ands target i compare current geometry and target shape, re- spectively. The weightλbalances the two.φ(I)is the normalized [CLS] token from a frozen DINOv3 ViT-B/16 backbone [ 22]. 4.1.3. VLM Invocation The Stage 1 VLM input–output mapping is X (1) t =VLM ( I ∗ , I tag t , T geom t , I ret t , o ret i ,P (1) ) The promptP (1) explains the inputs and requests the next command type, references, and parameters. The output follows the JSON schema in Table 1. 10 Extrude_add 퐼 푖 depth 퐼 푖 ∗ 퐼 푖 ret 표 푖 ret QueryResult Figure 8:A modeling example used for retrieval.Each record isI depth i , I ∗ i , I ret i , o ret i . I depth i andI ∗ i are retrieval keys; the retrieved information comprises a pseudo-previewI ret i and operation typeo ret i . In the pseudo-preview, translucent yellow marks the sketch plane, and translucent cyan marks other references, such as axes or paths. Red and blue outlines indicate the driving and terminal profiles; green indicates forming edges. Existing geometry is shown with depth shading and magenta edges. 4.2. Stage 2: Sketch Generation 4.2.1. VLM Invocation Before invoking the VLM, Stage 2 orients the camera normal to the se- lected sketch plane. The pseudo-bounding sphereS t (constructed under the registered Stage 1 camera) determines the translation and scale to help keep the intended sketch within the viewport ( Figure 6). The system then renders the tag imageI tag,s t and exports the geometry tableT geom,s t from this sketch view. The Stage 2 VLM input–output mapping is X (2) t =VLM [ I ∗ , I tag,s t , T geom,s t , X (1) t ,P (2) ] The promptP (2) requests one or more closed profiles in the view coordi- nates ofI tag,s t . The CAD kernel converts these coordinates into the sketch plane’s actual local coordinates. Projected edge samples inT geom,s t guide primitive placement. For example, a sketch line can use the projected end- points of an existing edge to align with it. The outputX (2) t defines primitives in JSON: a line uses two endpoints; an arc uses start, intermediate, and end points. A circle uses four ordered points on its circumference. 4.2.2. Sketch Constraint Generation Constraints are added in three stages: links to projected external ge- ometry, relationships among sketch primitives, and dimensions. Candidate relationships are detected from the predicted geometry using tolerance tests. 11 Table 1:Main fields of the Stage 1 outputX (1) t .TheTYPEfield accepts 16 val- ues: extrude_add, extrude_cut, symmetric_extrude_add, symmetric_extrude_cut, re- volve_add, revolve_cut, symmetric_revolve_add, symmetric_revolve_cut, sweep_add, sweep_cut, loft_add, loft_cut, fillet, chamfer, cplane_offset, and END. FieldDescription TYPEType of the next FreeCAD operation, or the termi- nation tokenEND. REQUIRE_SKETCHWhether Stage 2 must generate a new sketch. Stage 2 is skipped if this value isfalse. SKETCH_REFID of the sketch plane for a new sketch. SKETCH_HINTBrief description of the intended shape and loca- tion, guiding Stage 2. OTHER_REFERENCEOther reference IDs: directions, revolution axes, sweep paths, loft sections, fillet/chamfer edges, and construction-plane source faces. PARAMETER_HINTApproximate depth, angle, offset, fillet radius, or chamfer distance, relative to the view scale inI tag t . DIR_HINTQualitative direction in the current view, such as extrusion along the positive or negative side of the sketch plane. (1) Constraints to projected geometry.Existing edges are imported as external sketch geometry using their source IDs and projected coordinates. We apply constraints in the following order: (i) collinearity between a sketch line and a projected straight edge; (i) coincidence of a sketch line or arc endpoint with a projected edge endpoint or the projected center of an external circle or arc; and (i) point-on-object constraints that place a sketch line or arc endpoint on a projected edge. These constraints link the sketch to the source geometry. (2) Constraints among sketch primitives.We next apply endpoint co- incidence, horizontal and vertical line constraints, tangency involving circles or arcs, equal line lengths, equal arc radii, and point symmetry, in that or- der. Each candidate is tested using the positional, angular, length, or radius adjustment needed to satisfy the relationship. For symmetric constraint, we project the outer loop of the sketch face. If the projected outline is symmetric about both its horizontal and vertical midlines, their intersection and the two midlines define the reference center 12 and symmetry axes. Otherwise, we use the sketch origin and itsX/Yaxes. The same fallback applies when no supporting face is available. Symmetry constraints are detected on primitive endpoints and circle or arc centers. (3) dimensional constraints.Dimensional constraints specify line lengths, circle or arc radii, and key-point offsets from the reference center measured along the reference axes. Lengths and radii are measured after geometric constraints have been applied. FreeCAD solves the sketch after each constraint is added. Candidates that overconstrain the sketch are reverted, and constraint addition stops when no degrees of freedom remain. 4.3. Stage 3: Execution Verification FreeCAD executes the command defined byX (1) t and, when required,X (2) t to render a modeling previewI pv t ( Figure 9). The Stage 3 VLM input–output mapping is X (3) t =VLM [ X (1) t , I ∗ , I pv t ,P (3) ] The promptP (3) asks whether the candidate creates a missing target structure or a useful intermediate feature. The JSON outputX (3) t contains an acceptance decision and a brief acceptance justification or detailed rejection reasons. Accepted commands updateD t toD t+1 . Rejection reasons are added to the context before restarting Stage 1. Figure 9:Modeling previews used for execution verification. 5. Experiments 5.1. Geometry Explicit Reference Dataset (GERD) GERD contains 2,700 publicly available Onshape parts, split into 2,500 training, 100 validation, and 100 test samples. It covers extrusion, revolu- tion, sweep, loft, chamfer, fillet, and create construction plane. Each model- ing step records the command, resulting geometry, and available entity IDs 13 (Figure 10). To the best of our knowledge, GERD is the first dataset to align these records across the full Body, Face, Edge, and Vertex hierarchy at every step. "MainRef": ..., "Geometry": [..., ...], "Command": [..., ...] “MainRef": "JDC": "coord_sys": ... , "JCC": ..., "JEC": ... “Geometry": [ "Body": ..., “Face": ..., “Edge": ..., “Vertex": ... , ..., ... ] "Body": "JHD": "face_ids": ["JHC", "JHG", ...], "volume": 25.2 , "JGD": ..., ... "Face": "JHC": "surface_type": "cylinder", "edge_ids": ["JHB", "JHF", ...], "area": 7.2, "approx_bspline_face": "bspline_surface": ..., "outerloop_pcurve": ..., "innerloop_pcurves": [..., ..., ...], , "JHW": ..., ... "Edge": "JFB": "curve_type": "line", "vertex_ids": ["JYE", "JYs"], "length": 7.2, "approx_bspline_curve": ... , "JHW": ..., ... “Vertex": "JYo": [-3.6, 1.75, 0.5], "JYs": [...], ... "Command": [ "cmd_type": "extrude", "construct_type": "ADD", "region": ["JGC"], "depth": 1.0, "direction": null, "draft_angle": null, "is_symmetric": false, "is_opposite_dir": false, "id": "F0cE8lMPGTgcETL_0", "is_suppressed": false , ..., ... ] Figure 10:GERD data structure.Each record containsMainRef,Geometry, and Command.MainRefstores three main reference planes and their coordinate systems:Zis normal to the plane, andX/Ydefine local coordinates.Geometryindexes four levels by ID. Bodies store boundary face IDs and volume. Faces store surface type, boundary edge IDs, and area. Edges store curve type, endpoint vertex IDs, and length. Vertices store 3D coordinates. Faces and edges also provide approximate B-spline representations for reconstruction across CAD systems.MainRef,Geometry, andCommandshare the same ID namespace. GERD-RET uses the GERD training split for modeling example retrieval (Figure 8). GERD-EVL is derived from GERD for image-to-CAD code gen- eration and inherits its training, validation, and test splits. Each sample isI i , C fs i , M step i , M ex10p i . Here,I i is the completed-part image andC fs i is its cadfs program [ 23]. The two STEP filesM step i , M ex10p i store the original geometry and the result after increasing all extrusion depths by10%. GERD also supports the derivation of datasets with different viewpoints, 14 rendering modalities, and supervision targets. Potential tasks include next- command prediction, face/edge/vertex selection, B-Rep encoding, and vali- dating references across modeling steps. GERD thus serves both as a data source for this work and as a reusable intermediate representation for step- wise parametric CAD modeling. 5.2. Implementation Details All stages usegpt-5.6-solas VLM, with at most two failed verification attempts in Stage 3. Stage 1 retrieves up to three examples withλ= 0.6. We use FreeCAD 1.1 and render images at500×500pixels. Generation stops atENDor 25 modeling steps and returns the current document, which may represent an incomplete model. Training-based baselines are trained on the GERD-EVL training split to generate cadfs code [23]. Other baselines generate FreeCAD or CadQuery code and use thegpt-5.6-solVLM unless specified otherwise. Evaluation on DeepCAD uses the 100-part CADCoder-test100 subset because evaluating the full test set is computationally costly [ 10]. 5.3. Evaluation Metrics We report compilation success rate (CR) (For Vision2CAD, the generated FreeCAD document is marked as compiled successfully if it is non-empty and reports no errors), mean intersection over union (mIoU), Chamfer dis- tance (CD), Hausdorff distance (HD), and 95th-percentile Hausdorff distance (HD95). HD95 excludes the largest5% of pointwise distances. Geometric metrics are computed only for samples counted as successful under the CR criterion. These metrics describe the quality of the returned geometry and must be interpreted alongside CR. We evaluate static reconstruction and a specific parameter-editing proto- col. For editing, all additive and subtractive extrusion depths are simulta- neously multiplied by1.1in each generated program and its ground-truth counterpart. We then recompute CR and the geometric metrics against the edited ground truth. The motivation for conducting parameter-editing experiment is as fol- lows: a parametric CAD part can often be constructed through multiple valid feature histories, with different feature decompositions, reference enti- ties, and sketch constraints producing the same final geometry. Consequently, directly measuring reference or constraint accuracy against a single ground- truth construction history is not generally meaningful. Nevertheless, part of 15 the modeling intent is observable from the target image through geometric relationships. For example, the inner and outer cylindrical surfaces of an annular feature imply that the corresponding sketch circles should typically be concentric. To evaluate whether such image-supported design intent is captured by the generated parametric dependencies, we perturb the extru- sion depth by10% and examine how the model geometry responds to the parameter change. A model whose references and constraints correctly en- code these relationships is expected to preserve the corresponding geometric dependencies during editing. Therefore, this experiment provides an indi- rect assessment of whether the reconstructed model captures design intent through its parametric dependency structure. Although a single extrusion- depth perturbation cannot fully characterize model editability or complete design-intent preservation, it evaluates an important aspect of parametric behavior that cannot be measured static model reconstruction. 5.4. Experimental Results Static reconstruction.Vision2CAD returns non-empty documents with no reported errors for all test cases (100% CR) and achieves the best reported geometric metrics on both datasets (Tables 2and3). On GERD-EVL, mIoU improves by11.1% and CD decreases by17.3% relative to the strongest baseline for each metric. These comparisons are based on each method’s outputs that satisfy both CR conditions: CAD-Coder’s mIoU of0.4261covers only 18 of the 100 test parts, whereas Vision2CAD’s scores cover all 100. DeepCAD mainly contains regular sketch–extrusion sequences. Most baselines reach at least97% CR, so the evaluated subsets differ less in size than on GERD-EVL. Vision2CAD improves mIoU by5.6% and reduces CD by41.8% relative to the strongest baseline for each metric. These results indicate improved geometric reconstruction on the evaluated outputs. Parameter editing.Vision2CAD retains100% CR and the best reported geometric metrics after the uniform extrusion-depth increase ( Tables 4and5). Relative to the strongest baseline for each metric, mIoU improves by7.0% on GERD-EVL and4.2% on DeepCAD. CD decreases by20.2% and42.0%, respectively. All Vision2CAD outputs are more closely matches the edited ground truth according to the reported scores. Qualitative results.In the GERD-EVL examples in Figure 11, baseline outputs miss holes, supports, or curved profiles, and some methods return no geometry. Vision2CAD recovers these structures more consistently in the illustrated cases. Most methods recover the basic shapes of the DeepCAD 16 Table 2:Static geometric reconstruction results on GERD-EVL. CategoryMethodCR (%)↑mIoU↑CD↓HD↓HD95↓ Trained models ReCAD-SFT43.0 0.2931 0.1103 0.2004 0.1430 ReCAD-GRPO44.0 0.2980 0.0955 0.1912 0.1211 CAD-Coder18.0 0.4261 0.0832 0.1596 0.1098 Training-free agents CADCodeVerify89.0 0.3989 0.1017 0.1975 0.1297 CAD-Assistant100.0 0.4038 0.0721 0.1558 0.0909 Direct generation gpt-5.6-sol84.0 0.3920 0.0964 0.1822 0.1139 deepseek-v4-pro75.0 0.3570 0.0922 0.1741 0.1108 OursVision2CAD100.0 0.4733 0.0596 0.1125 0.0663 Table 3:Static geometric reconstruction results on the DeepCAD test sub- set [10]. CategoryMethodCR (%)↑mIoU↑CD↓HD↓HD95↓ Trained models ReCAD-SFT98.0 0.6421 0.0743 0.1289 0.0889 ReCAD-GRPO99.0 0.6051 0.0762 0.1361 0.0891 CAD-Coder97.0 0.6619 0.0637 0.1270 0.0847 Training-free agents CADCodeVerify99.0 0.6038 0.0586 0.1065 0.0676 CAD-Assistant100.0 0.6170 0.0631 0.1125 0.0743 Direct generation gpt-5.6-sol98.0 0.5988 0.0550 0.0949 0.0630 deepseek-v4-pro96.0 0.4636 0.0951 0.1560 0.1115 OursVision2CAD100.0 0.6987 0.0320 0.0551 0.0332 examples, while Vision2CAD better matches their local contours and feature positions. 5.5. Ablation Studies We evaluate six ablations on GERD-EVL under the same static and edit- ing protocols. 1.w/o EGR (explicit geometry referencing):Stage 1 removes edge ID labels and leader lines fromI tag t and omitsT geom t . CadQuery selec- tors replace explicit entity IDs (e.g.,>Zselects the highest face and|Z selects edges parallel toZ). 2.w/o LCB (local-coordinate bridge):Stage 2 omits the bridge from view coordinates to sketch local coordinates. The VLM must instead 17 Table 4:Parameter-editing results on GERD-EVL after increasing all additive and subtractive extrusion depths by 10%. CategoryMethodCR (%)↑mIoU↑CD↓HD↓HD95↓ Trained models ReCAD-SFT43.0 0.2961 0.1096 0.1972 0.1406 ReCAD-GRPO44.0 0.2957 0.0969 0.1925 0.1203 CAD-Coder15.0 0.4458 0.0883 0.1662 0.1176 Training-free agents CADCodeVerify89.0 0.3967 0.1027 0.1984 0.1311 CAD-Assistant98.0 0.4055 0.0759 0.1608 0.0937 Direct generation gpt-5.6-sol84.0 0.3906 0.0977 0.1826 0.1160 deepseek-v4-pro75.0 0.3620 0.0930 0.1755 0.1115 OursVision2CAD100.0 0.4770 0.0606 0.1149 0.0687 Table 5:Parameter-editing results on the DeepCAD test subset [10] after in- creasing all additive and subtractive extrusion depths by 10%. CategoryMethodCR (%)↑mIoU↑CD↓HD↓HD95↓ Trained models ReCAD-SFT98.0 0.6346 0.0739 0.1276 0.0873 ReCAD-GRPO99.0 0.6035 0.0771 0.1360 0.0900 CAD-Coder97.0 0.6595 0.0643 0.1276 0.0845 Training-free agents CADCodeVerify99.0 0.5950 0.0623 0.1096 0.0707 CAD-Assistant100.0 0.5950 0.0676 0.1159 0.0769 Direct generation gpt-5.6-sol98.0 0.5860 0.0591 0.0979 0.0667 deepseek-v4-pro96.0 0.4747 0.0954 0.1548 0.1120 OursVision2CAD100.0 0.6875 0.0343 0.0581 0.0362 predict sketch primitive coordinates directly in the sketch plane’s actual local coordinate system. 3.w/o PGR (projected geometry referencing):Stage 2 removes edgesamplesfromT geom,s t and disables sketch constraints to projected external geometry. 4.w/o Reg (registration):Stage 1 disables image registration and the associated sphereS t . Stage 2 removes the[0,1]restriction on predicted sketch coordinates. 5.w/o RAG (retrieval):Stage 1 does not retrieve similar modeling examples from GERD-RET. The VLM predicts the next command using only the current model state and target image. 6.w/o EVf (execution verification):Remove Stage 3 (the preview- 18 ReCAD -GRPO Compile Failed Compile Failed Compile Failed CAD -Coder Compile Failed Compile Failed Compile Failed CAD -Assistant deepseek -v4-pro Compile Failed gpt -5.6-sol Compile Failed CAD CodeVerify Compile Failed ReCAD -SFT Compile Failed Compile Failed Compile Failed vision2cadImage Figure 11:Reconstruction results on GERD-EVL and DeepCAD.The top four rows show GERD-EVL examples; the bottom four rows show DeepCAD examples. based verification). Modeling commands generated by Stages 1 and 2 are committed without visual verification. All configurations achieve100.0% CR in static reconstruction, but the full system gives the best geometric scores (Table 6). Editing exposes larger differences in execution success ( Table 7). Removing explicit geometry referencing (EGR) causes the largest static mIoU drop, from0.4733to0.4187. After editing, CR falls to98.0% and mIoU to0.4128, the lowest among the ablations. This supports the importance of accurate entity selection throughout the feature sequence. Without the local-coordinate bridge (LCB), static mIoU falls to0.4332 and editing CR to90.0%. Sketch coordinate errors thus affect both initial geometry and later updates. Deterministic coordinate conversion reduces this source of error. Projected geometry referencing (PGR) has a modest static effect: mIoU falls only to0.4658. After editing, however, CR falls to89.0% and mIoU to 19 Table 6:Ablation results for static reconstruction on GERD-EVL.CR counts non-empty documents with no reported errors; geometric metrics are conditional on those outputs. Configuration CR (%)↑mIoU↑CD↓HD↓HD95↓ Full100.0 0.4733 0.0596 0.1125 0.0663 w/o EGR100.0 0.4187 0.0891 0.1852 0.1143 w/o LCB100.0 0.4332 0.0852 0.1726 0.1045 w/o PGR100.0 0.4658 0.0639 0.1133 0.0938 w/o Reg100.0 0.4589 0.0623 0.1374 0.0889 w/o RAG100.0 0.4522 0.0649 0.1241 0.0785 w/o EVf100.0 0.4478 0.0724 0.1451 0.0924 0.4263. Predicting nearby sketch coordinates can reproduce an initial profile, but cannot replace persistent references to source geometry. PGR provides those references and supports subsequent parameter propagation. Registration, RAG, and execution verification also improve geometric scores. Removing them reduces editing CR to96.0%,98.0%, and95.0%, respectively. Their contributions complement the reference mechanisms through visual alignment, modeling examples, and rejection of incorrect operations. 6. Conclusions We presented Vision2CAD, a visual agent harness for generating native parametric CAD models from single images. It combines explicit entity refer- ences, a local-coordinate bridge, and sketch constraints to projected external geometry. These mechanisms establish feature dependencies within the sup- ported modeling operations and constraint types. GERD provides aligned commands, geometric states, and entity IDs for research on these dependen- cies. Vision2CAD achieves the best geometric accuracy among the evaluated methods on GERD-EVL and the DeepCAD test subset. It also retains a 100% compilation success rate after extrude depths increase by10%. Abla- tion results highlight the role of explicit dependencies in preserving editabil- ity, especially when static geometry alone appears accurate. 20 Table 7:Ablation results on GERD-EVL after increasing all additive and sub- tractive extrusion depths by 10%.CR counts non-empty documents with no reported errors; geometric metrics are conditional on those outputs. Configuration CR (%)↑mIoU↑CD↓HD↓HD95↓ Full100.0 0.4770 0.0606 0.1149 0.0687 w/o EGR98.0 0.4128 0.0875 0.1590 0.0905 w/o LCB90.0 0.4139 0.1035 0.1826 0.1380 w/o PGR890.4263 0.0882 0.1487 0.1029 w/o Reg96.0 0.4483 0.0817 0.1573 0.1024 w/o RAG98.0 0.4585 0.0758 0.1332 0.0827 w/o EVf95.0 0.4367 0.0810 0.1516 0.1029 The system remains limited by its single target view and dense edge labels on complex parts. It does not yet support references to entities internal to sketches or operations such as linear and circular patterns. Future work will extend these capabilities and explore multiple views, active viewpoint planning, and adaptive entity visualization. CRediT authorship contribution statement First Author: Conceptualization, Methodology, Software, Validation, Writing – original draft. Second Author: Supervision, Writing – review and editing. Declaration of competing interest The authors declare that they have no known competing financial inter- ests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. 21 Data availability The data and code supporting the findings of this study will be made available at [repository and persistent identifier] upon publication. References [1]C.-G. Yin, Y.-S. Ma,Parametric feature constraint modeling and map- ping in product development, Advanced Engineering Informatics 26 (3) (2012) 539–552.doi:10.1016/j.aei.2012.02.010. URLhttps://doi.org/10.1016/j.aei.2012.02.010 [2]Y. Cheng, F. He, X. Lv, W. Cai,On the role of generating textual description for design intent communication in feature-based 3D collab- orative design, Advanced Engineering Informatics 39 (2019) 331–346. doi:10.1016/j.aei.2019.02.003. URLhttps://doi.org/10.1016/j.aei.2019.02.003 [3]T. Chen, C. Yu, Y. Hu, J. Li, T. Xu, R. Cao, L. Zhu, Y. Zang, Y. Zhang, Z. Li, L. Sun,Img2CAD: Conditioned 3-d CAD model generation from single image with structured visual geometry, IEEE Transactions on Industrial Informatics 21 (11) (2025) 8539–8549.doi: 10.1109/TII.2025.3584476 . URLhttps://doi.org/10.1109/TII.2025.3584476 [4]C. Chen, J. Wei, T. Chen, C. Zhang, X. Yang, S. Zhang, B. Yang, C.-S. Foo, G. Lin, Q. Huang, F. Liu,CADCrafter: Generating computer- aided design models from unconstrained images, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, p. 11073–11082. URL https://openaccess.thecvf.com/content/CVPR2025/html/ Chen_CADCrafter_Generating_Computer-Aided_Design_Models_ from_Unconstrained_Images_CVPR_2025_paper.html [5]R. Wu, C. Xiao, C. Zheng,DeepCAD: A deep generative network for computer-aided design models, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, p. 6772–6782. URL https://openaccess.thecvf.com/content/ICCV2021/html/ 22 Wu_DeepCAD_A_Deep_Generative_Network_for_Computer-Aided_ Design_Models_ICCV_2021_paper.html [6]M. S. Khan, S. Sinha, T. U. Sheikh, D. Stricker, S. A. Ali, M. Z. Afzal,Text2CAD: Generating sequential CAD designs from beginner- to-expert level text prompts, in: Advances in Neural Information Processing Systems, Vol. 37, Curran Associates, Inc., 2024, p. 7552– 7579.doi:10.52202/079017-0242. URLhttps://proceedings.neurips.c/paper_files/paper/2024/ hash/0e5b96f97c1813b75f6c28532c2ecc7-Abstract-Conference. html [7]W. Ma, S. Chen, Y. Lou, X. Li, X. Zhou,Draw step by step: Recon- structing CAD construction sequences from point clouds via multimodal diffusion , in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, p. 27154–27163. URL https://openaccess.thecvf.com/content/CVPR2024/html/ Ma_Draw_Step_by_Step_Reconstructing_CAD_Construction_ Sequences_from_Point_CVPR_2024_paper.html [8]D. Rukhovich, E. Dupont, D. Mallis, K. Cherenkova, A. Kacem, D. Aouada,CAD-Recode: Reverse engineering CAD code from point clouds , in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, p. 9801–9811. URLhttps://openaccess.thecvf.com/content/ICCV2025/html/ Rukhovich_CAD-Recode_Reverse_Engineering_CAD_Code_from_ Point_Clouds_ICCV_2025_paper.html [9]A. Joglekar, A. Regmi, K. Shimada, L. B. Kara,Ortho2CAD: 3d CAD generation from orthographic drawings using vision language models, arXiv preprint arXiv:2607.08891 (2026).arXiv:2607.08891. URL https://arxiv.org/abs/2607.08891 [10]A. C. Doris, M. F. Alam, A. H. Nobari, F. Ahmed,CAD-Coder: An open-source vision-language model for computer-aided design code gen- eration , arXiv preprint arXiv:2505.14646 (2025).arXiv:2505.14646. URLhttps://arxiv.org/abs/2505.14646 [11]D. Mallis, A. S. Karadeniz, S. Cavada, D. Rukhovich, N. Foteinopoulou, K. Cherenkova, A. Kacem, D. Aouada,CAD-Assistant: Tool-augmented 23 VLLMs as generic CAD task solvers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, p. 7284–7294. URLhttps://openaccess.thecvf.com/content/ICCV2025/html/ Mallis_CAD-Assistant_Tool-Augmented_VLLMs_as_Generic_CAD_ Task_Solvers_ICCV_2025_paper.html [12]X. Li, J. Li, Y. Song, Y. Lou, X. Zhou,Seek-CAD: A self-refined genera- tive modeling for 3d parametric CAD using local inference via DeepSeek, in: International Conference on Learning Representations (ICLR), 2026. URLhttps://openreview.net/forum?id=PzIc2TxhwN [13]J. Barkley, R. Loghmani, A. B. Farimani,CADSmith: Multi-agent CAD generation with programmatic geometric validation, arXiv preprint arXiv:2603.26512 (2026).arXiv:2603.26512. URLhttps://arxiv.org/abs/2603.26512 [14]D. Qi, C. Wang, J. Xu, T. Chu, Z. Zhao, W. Liu, W. Ding, Y. Ma, S. Gao,Pointer-CAD: Unifying B-Rep and command sequences via pointer-based edges & faces selection , in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, p. 17377–17387. URLhttps://openaccess.thecvf.com/content/CVPR2026/html/ Qi_Pointer-CAD_Unifying_B-Rep_and_Command_Sequences_via_ Pointer-based_Edges__CVPR_2026_paper.html [15]M. Yuan, D. Lin, H. Xie, T. Wu, R. Ma,CAD-Refiner: A unified frame- work for CAD generation and iterative editing, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, p. 3244–3253. URL https://openaccess.thecvf.com/content/CVPR2026/html/ Yuan_CAD-Refiner_A_Unified_Framework_for_CAD_Generation_ and_Iterative_Editing_CVPR_2026_paper.html [16]J. Li, Y. Luo, Y. Lou, X. Zhou,ReCAD: Reinforcement learning en- hanced parametric CAD model generation with vision-language mod- els , in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, 2026, p. 6190–6198.doi:10.1609/aaai.v40i8.37544. URLhttps://doi.org/10.1609/aaai.v40i8.37544 24 [17]CadQuery Developers, Selectors reference,https://cadquery. readthedocs.io/en/latest/selectors.html, cadQuery Documenta- tion, accessed August 18, 2026 (2026). [18]X. Liu, Y. Zhang, L. Wang, L. Gu, Y. Xu, K. Zheng, Q. Cai, G. Zhou, LLMs driven fusion AI-AD system for mechanical design: From un- derstanding to generation, Advanced Engineering Informatics 68 (2025) 103745.doi:10.1016/j.aei.2025.103745. URLhttps://doi.org/10.1016/j.aei.2025.103745 [19]Z. Yuan, J. Shi, Y. Huang, Openecad: An efficient visual language model for editable 3d-cad design, Computers & Graphics 124 (2024) 104048. [20]S. Wang, C. Chen, X. Le, Q. Xu, L. Xu, Y. Zhang, J. Yang,CAD- GPT: Synthesising CAD construction sequence with spatial reasoning- enhanced multimodal LLMs, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, p. 7880–7888.doi:10.1609/ aaai.v39i8.32849 . URLhttps://doi.org/10.1609/aaai.v39i8.32849 [21]X. Xu, K. D. D. Willis, J. G. Lambourne, C.-Y. Cheng, P. K. Jayaraman, Y. Furukawa, SkexGen: Autoregressive generation of CAD construction sequences with disentangled codebooks, in: Proceedings of the 39th International Conference on Machine Learning, PMLR, 2022, p. 24698– 24724. URL https://proceedings.mlr.press/v162/xu22k.html [22]O. Siméoni, H. V. Vo, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V. Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haz- iza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. Jé- gou, P. Labatut, P. Bojanowski, DINOv3(2025).arXiv:2508.10104. URL https://arxiv.org/abs/2508.10104 [23]V. Pyatov, G. Bobrovskikh, S. Galochkin, N. Boldyrev, O. Voynov, A. Filippov, G. Ferrer, P. Wonka, E. Burnaev,CADFS: A big CAD program dataset and framework for computer-aided design with large language models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.arXiv: 25 2605.01925. URLhttps://arxiv.org/abs/2605.01925 26