The one-paragraph version
HistCAD is a new way to represent, share, and test parametric computer-aided design models that cares about editability, not just final shape. In real engineering, a CAD model is valuable because you can change one dimension — make this hole bigger, make that line longer — and the rest of the model updates sensibly because explicit rules hold things together. The authors build an intermediate language that records those rules explicitly, a dataset of 170,236 executable models that unifies academic sources with professionally authored industrial parts, and a benchmark that edits a dimension and checks whether the required rules still hold. Their core finding is stark: if you drop the rules and keep only enough information to close the sketch outlines, models still rebuild almost every time but lose design intent about half the time on DeepCAD-derived data. With full constraints and industrial data, text-to-CAD generation gets more distributionally complete and more editable, though industrial models remain the hardest test.
The problem they're solving, and why it matters
A parametric CAD model is not just a static three-dimensional shape. It is a construction history: a list of steps a CAD system can replay. A typical history starts with a two-dimensional sketch drawn on a plane, then turns that sketch into solid material by operations like extruding outward or cutting inward.
What makes that history reusable is design intent, meaning the relationships the designer meant to stay true no matter how dimensions change. Design intent is encoded as constraints. A dimensional constraint says how big something is, for example the length of a line is 20 millimeters or the diameter of a circle is 16 millimeters. A geometric constraint says how things relate, for example two lines stay perpendicular, two circles stay concentric, two edges stay tangent.
Existing generative CAD research largely ignored those constraints in both data and evaluation. DeepCAD gives large-scale histories of sketches plus extrusions but omits explicit sketch constraints. Text2CAD adds language descriptions on top of the same constraint-free format. SketchGraphs has rich two-dimensional constraints but no executable three-dimensional steps to build a solid. Fusion 360 Gallery has authentic human-authored designs with constraints in their native files, but its reconstructable subset is small, about 8,625 designs, and mostly limited to sketch plus extrude. Evaluation inherited the same blind spot: it measured shape similarity, whether code runs once, or whether predicted commands match tokens, but not whether a model still respects its required relationships after you change a parameter.
That matters because a model that looks right can break under the first edit. The paper's motivating example lengthens a line to 173 millimeters. Without explicit constraints, perpendicular corners shear and the shape distorts. With explicit constraints, the solver holds perpendicularity while honoring the new length. The authors call the missing capability parameter editability with constraint preservation: accept a local parameter change while keeping the relevant dimensional and geometric relationships satisfied.
The key idea, in plain words
Stop treating CAD generation as imitating a final shape. Treat it as synthesizing a reusable recipe where every important relationship is written down.
To do that, HistCAD separates the logical recipe from any particular CAD software. Think of it like a neutral script that any backend — currently JiuShao POWER Platform and Autodesk Fusion 360 — can read and rebuild as a native parametric model. The script lists flat sketch pieces, the full graph of constraints on those pieces, the solid-building steps, and a software-independent way to point at existing edges for operations like fillet, which rounds an edge, and chamfer, which bevels an edge.
Two simplifications make this practical for language models. First, sketches are stored as a flat unordered set of atomic pieces rather than a nested hierarchy of faces containing loops containing curves. Shared interior borders appear twice in hierarchical formats and are repeated as tokens; HistCAD cancels them out, keeping only the outer contour and hole borders that actually define the region to extrude. Second, edge selections are stored as three-dimensional reference points near the target edge rather than internal topology numbers that differ between CAD kernels, learned pointer vectors, or a separate grounding module. At rebuild time the backend finds the nearest edge to that cue.
On top of this representation the authors build scale plus realism: combine DeepCAD scale, SketchGraphs constraint richness, and Fusion 360 human authorship with 8,093 professionally authored industrial standard parts that use longer histories and operations academic data lacks. Then they test not static fidelity but what happens after an edit.
How it actually works, step by step
First, the representation. A HistCAD sketch has a plane defined by position and orientation, a set of geometric primitives, meaning basic drawing elements like lines, circles, arcs, ellipses, elliptical arcs, and NURBS curves, which are smooth flexible curves controlled by control points. Across the full dataset lines dominate at about 71 percent, followed by circles and arcs. The sketch also has an explicit constraint system covering 19 types, including coincident for sharing endpoints, horizontal, vertical, parallel, perpendicular, concentric, tangent, normal, length, distance, diameter, radius, angle, minor and major radius, fix for anchoring, midpoint, equal, and mirror. Feature operations, meaning steps that create or modify solid material, include extrude defined by direction and extent, revolve defined by axis and angle, helix sweep defined by axis, pitch, turns and handedness for threads, plus fillet and chamfer via reference points. Each feature carries a Boolean tag that says how it combines with existing material: new body for a separate solid, join for adding material, cut for removing material, or intersect for keeping only overlap.
Second, dataset construction. The academic portion aligns DeepCAD sequences with SketchGraphs constraints by matching shared sketch entities, splitting intersecting curves into disjoint atomic pieces, recovering the selected region by the symmetric-difference idea described above, merging collinear fragments, transferring original constraints, adding auxiliary constraints needed for execution, and pruning redundancies. That yields 153,534 sequences with explicit constraints, plus 8,609 recovered sketch-to-extrusion sequences from Fusion 360 Gallery, for 162,143 academic sequences. The industrial portion parses 8,093 native industrial models into the same format. Those models are longer on average and extend coverage: about 38 percent contain at least one fillet or chamfer, about 4 percent contain revolves, and about 2 percent contain helical sweeps, categories absent from the academic portion. Every sequence is aligned with native CAD files, STEP exports, which are a neutral exchange format for solid geometry, rendered views, and text annotations derived from the sequence itself using deterministic loop-nesting and oriented-bounding-box analysis followed by Gemma-4-31B-it to describe modeling process, geometric structure, and functional type.
Third, execution. A backend reads steps in order, creates the sketch coordinate system, instantiates primitives, inserts constraints, and calls native feature constructors. For fillets and chamfers it binds the stored point cue to the nearest rebuilt edge within tolerance. Adapters to SolidWorks and CATIA are listed as planned.
Fourth, the Constraint-Aware Editability Benchmark. Each test case starts from an executable source sequence, specifies a target dimensional edit to a length, radius, diameter, angle, or distance constraint, gives the concrete new value and where to attach it — either modifying an existing dimension or inserting a new one — and fixes a preservation set of other constraints that should still hold. The evaluated sequence receives the same edit. Three checks produce three rates explained in words first: Edit Reachability is whether the edited model satisfies the new target, passes sketch validation, and rebuilds downstream; Conditional Preserved Constraint Satisfaction Rate is, among those reachable cases, whether every constraint in the preservation set still holds; Overall Editable Success is whether both happen, mathematically the product of the first two. This separation distinguishes failing to get any valid edited solid from getting a solid that silently violates intent.
What they showed — results, honestly, with limitations
Representation compactness: on a shared intersection of 134,896 models, HistCAD with full constraints averages about 550 tokens with the Gemma tokenizer, compared with about 591 for Text2CAD and about 1,974 for DeepCAD. HistCAD without constraints is even shorter at about 364 tokens. The constraints cost roughly 186 extra tokens on average — modest for what they enable. Industrial sequences are much longer, with mean around 1,861 tokens and 95th percentile around 6,323.
Native ablation proves constraints matter for preservation, not just rebuild. Keeping full constraints gives about 95 percent Edit Reachability and 100 percent conditional preservation in their validation, hence about 95 percent overall success. Keeping only coincidence constraints needed for closed outlines — the closure-only variant — still reaches about 96 percent reachability but drops to about 68 percent conditional preservation and about 65 percent overall. On the DeepCAD-derived subset the gap is sharper: closure-only still reaches 99.5 percent but preserves only 55.5 percent, so nearly half of reachable edits violate something. By constraint type, equal relations are most fragile at about 60 percent preserved under closure-only, followed by angle, parallel, length, and perpendicular; coincident stays at 100 percent because it is explicitly retained. Adding a new dimension is harder than modifying an existing one mainly because preservation drops, not reachability.
Text-to-CAD generation with Qwen3-8B plus LoRA, a parameter-efficient fine-tuning method, shows full HistCAD training gives the best aggregate shape and editability: best median shape error, coverage, and distribution match, plus highest constraint prediction F1 around 62 percent, reachability around 84 percent, conditional preservation around 87 percent, and overall editable success around 73 percent. Adding only Fusion 360 data improves constraint F1 but not overall editable success; the industrial data is needed for the strongest overall result. By subset, the DeepCAD-only baseline wins on DeepCAD test data due to distributional proximity, but full HistCAD wins clearly on industrial data and improves coverage there from about 52 to about 56 percent. Industrial remains hardest: even the best model reaches only 64 percent overall editable success there versus 85 percent on Fusion 360, failing both on reachability and on preservation.
A qualitative study prompts GPT-5.5 with only the schema and a natural-language request and gets executable JSON for a filleted flange with bolt holes, a hex nut with bore and chamfers, plus edited variants adding a raised neck and a helical thread cut, all rebuilt by the adapters. This suggests the schema is directly generatable by general models, but it is a case study, not a measured benchmark.
Limitations are stated plainly: the benchmark currently covers only local dimensional edits, not multi-step edits, feature-level changes, or reordering; the industrial subset at about 8,000 models is small relative to the academic portion, limiting signal for complex operations; text annotations are model-generated rather than designer-authored, so their quality as supervision needs further study; and cross-kernel validation beyond the two current backends remains future work. It is also unclear how sensitive the point-cue binding for fillets and chamfers is to tolerance and geometry changes.
Why this might matter to me
For spatial reasoning in LLMs and VLMs, this reframes correctness from looking right to staying right under intervention. The benchmark's split between reaching a valid state and preserving relations is a useful template for testing whether a model truly encodes perpendicularity, concentricity, or symmetry rather than memorizing coordinates. The per-type fragility ranking — equal and angle relations break first — gives concrete targets for constraint-aware prompting or verification.
For world models, HistCAD offers histories where actions have solver-mediated consequences. Predicting the next token is not enough; predicting how a dimension change propagates through a constraint graph is closer to a physics of designed objects. That propagation could serve as supervision or as a rollout check for imagined edits.
For text-to-CAD and parametric geometry, the neutral flat-plus-constraints language plus point-based edge references is immediately reusable: it is compact for limited context windows, executable without kernel-specific IDs, and directly generatable as JSON by off-the-shelf LLMs. The finding that closure-only models rebuild but violate intent warns against training text-to-CAD on geometry alone.
For 3D generation and scene representation, the lesson is representational: separate topology-defining relations from numeric values, store selections as geometric cues rather than indices, and evaluate with counterfactual edits. The same pattern could extend from single parts to assemblies and scenes where objects must stay attached, aligned, or proportioned when parameters change.
Terms worth knowing
Parametric CAD: design by parameters and rules that a solver maintains, so changing a number updates the model rather than redrawing it.
Construction history: the ordered replayable steps — sketches plus solid operations — that build a model.
Sketch primitive: an atomic two-dimensional element such as a line, circle, arc, ellipse, or NURBS curve.
Constraint, dimensional versus geometric: a dimensional constraint sets size such as length or radius; a geometric constraint sets relationships such as parallel, perpendicular, coincident, tangent, concentric, equal, or mirror.
Design intent: the relationships meant to persist through edits; what the benchmark tries to verify.
Feature operation: a solid-building step such as extrude, revolve, sweep, fillet, chamfer, combined by Boolean operations join, cut, intersect, or new body.
B-Rep: boundary representation, the CAD kernel's description of a solid by faces, edges, and vertices and how they connect.
Flat sketch representation: storing only atomic border pieces without explicitly listing which face and loop each belongs to; shared interiors cancel out, outer and hole borders remain.
STEP file: a standard neutral file for exchanging solid geometry between CAD systems.
Edit Reachability, Conditional Preserved Constraint Satisfaction Rate, Overall Editable Success: respectively, did the edit produce a valid rebuilt model meeting the new target; among those, did all required old constraints still hold; and did both happen.