acceptodds
Under review as a conference paper at ICLR 2027

IntentCAD: Learning Text-to-CAD from the Intent Behind the Shape

Abstract

A CAD program that reproduces the right shape can still hide how the part was designed. The same mounting plate can be written as four separate circle cuts at literal coordinates or as one hole pattern with a shared diameter, and many text-to-CAD training pairs take the first form, with descriptions that leave the part’s purpose unstated. We ask whether making this design information explicit in training improves generation even when test prompts describe geometry alone. To test this, we construct IntentCAD, 446K text–program pairs from 223K parts, whose programs organize operations into features with named parameters and explicit relations, are verified against the source geometry, and are paired with descriptions that also state the part’s use and design intent. In controlled experiments with a 4B model on Text2CAD-Bench, normalizing programs and descriptions lowers Chamfer distance by 27% on average. Adding design intent, which no test prompt mentions, lowers it by a further 11% and raises IoU on all four splits. The resulting model also produces valid programs more often than four general- purpose LLMs, with an average invalid rate of 11.0% against 20.7–63.6%. These results suggest that what a text-to-CAD model learns depends on the design information its training data exposes, beyond the geometry it reproduces.

Primary area:
generative models
Keywords:
LLMs, Parametric-CAD, Text-to-CAD, Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9i8aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an analysis to isolate the effect of disclosing the "intent" behind a CAD design in a CAD prompt. To do so they build a dataset with 223k CAD shapes labeled by gpt-5.2, requiring it to include the "intent" in the prompt, and compare the finetuning of a 4B LLM on such data vs data without "intent" and claim to observe that the one fine tuned with the intent actually achieves better accuracy on test data with no intent in the prompts, and that the resulting model has a lower invalidity ratio than four general-purpose LLMs.

Critical strengths

The paper poses a reasonable and interesting hypothesis: that adding intent in the training data of a generative CAD model is a substantial information addition that can improve CAD generation capabilities, even when test prompts do not contain it. Characterizing the properties that improve the quality of CAD training data is an effective strategy to improve CAD models' generative power. If the hypothesis is confirmed empirically this would represent a meaningful contribution to the AI for CAD literature.

Critical weaknesses

The paper does not isolate what it claims to isolate.

  • "Program normalization" and "description normalization" are never defined and it is hard to understand what their method is comparing to. Also, it is not clear how the unnormalized programs and descriptions are built for shapes without a construction history.
  • It is not clear how the descriptions without intent are built, since the prompts in the Appendix always generate the intent paragraphs.
  • Confounders are not isolated: the proposed re-annotation changes several things together and the claimed gains might come from the annotator, the length of the annotations, the fact that there are more geometrical details, or the fact that the descriptions copy the style of the Text2CAD-Bench prompts. Indeed, by construction of the prompt, the "design intent" paragraph often restates geometry.

This is important because the effect of the intent is the main claim of the paper, and the current setup does not allow attributing the gains to it.

Anyways, the empirical evidence is very weak.

  • No variance or confidence intervals, while some gains are very tiny and possibly not significant.
  • A usual failure mode in CAD evals is that metrics are computed and averaged over valid outputs, which might surface a bias: a variant can look better if it fails on harder parts. This happens here too, e.g., adding intent raises the invalidity ratio on L1 Seq while lowering CD.
  • I am afraid some 3D shapes might live in the intersection between the training and test data, and no deduplication against Text2CAD-Bench is reported.
  • The evaluation setup employed in Table 3 for the specialized text-to-CAD models is very different from the one used to evaluate IntentCAD, since their numbers are copied from Wang et al.. Also, the LLMs are allowed up to three retries and it is not clear if IntentCAD is too.
  • There is no comparison with the same model fine-tuned on existing datasets (e.g., CADFS or Text-to-CadQuery) with a matched budget.
  • Results are restricted to one single model and to the two easiest tiers of Text2CAD-Bench.
  • In general, the claimed wins are far from being evident and are, at least, overstated: even under their own protocol, IntentCAD ranks 2nd to 4th on CD and IoU on 3 of 4 splits.

This is important because, even setting aside the first point, the results do not support the claims in the abstract.

Questions for authors

My main concerns require new experiments and a substantial rewrite, so I do not expect a rebuttal to be enough to change my recommendation. Still, the following questions could affect my assessment.

Clarifications

  • Are you positive you are not training on shapes present in Text2CAD-Bench? Did you run any deduplication (exact or near-duplicate) between your training sources and the benchmark?
  • How are the descriptions without intent built? With a separate prompt, or by dropping the first two paragraphs of the descriptions with intent?
  • How are the unnormalized programs and descriptions built for shapes without a construction history (e.g., ABC or the industrial parts)?
  • Some invalidity ratios are not multiples of 0.5, which they should be with 200 prompts per split: the unnormalized row in Table 2 and most LLM rows in Table 3. How were they computed?
  • Does IntentCAD also get up to three retries, like the LLMs in Table 3? Which decoding settings do you use?
  • Regarding the human reviews, how many regeneration rounds did you perform, how large was each sample, and what was the error rate?
  • Why only the L1 and L2 tiers of Text2CAD-Bench? L3 also includes parts with complex topology, which seems in scope for your claims.
  • Will the dataset be released, and how many of the 223K parts come from private sources?

Missing experiments

These are, in my opinion, the minimum required to support the claims of the paper:

  • Ablations to isolate the impact of including "intent" in the prompt, e.g., replacing the intent paragraph with a neutral or paraphrased paragraph of the same length.
  • Confidence intervals over multiple seeds, and metrics that count invalid outputs as failures instead of averaging over valid outputs only.
  • A comparison with the same model fine-tuned on existing datasets with a comparable training budget.
Reviewer_F3f2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IntentCAD, a framework for text-to-CAD generation that explicitly incorporates design information into the training data. Using existing CAD datasets and industrial sources, the authors construct a dataset containing approximately 446K text-program pairs, with descriptions that include the purpose and design intent of parts. Experiments on Text2CAD-Bench compare different levels of program and description normalization, as well as the inclusion of design intent.

Critical strengths
  1. The motivation is clear. The authors observes that existing text-to-CAD training pairs often describe final geometry while omitting feature organization, parameter relations, and part purpose. To address this, the authors proposed IntentCAD, which incorporates standardized CAD programs into the data and supplements it with descriptions of the component's purpose and design intent.
  2. The method has a clear framework. First, CAD programs are standardized to reveal their characteristic structures and the interdependencies among their parameters. Next, the descriptive information for each pair is augmented with details regarding the part’s intended use and design intent. Finally, a language model is trained using these enhanced "text-program" pairs.
Critical weaknesses
  1. Limited innovation. CADFS, HistCAD, CADmium, and related studies already use explicit parameters or constraints, construction-level descriptions, and functional annotations. IntentCAD appears primarily to recombine these ideas. The paper does not provide a component-level comparison or a fair experiment against the closest methods.
  2. The causal effect of “design intent” is not established. Intent labels are generated by GPT-5.2 from geometry, views, and reconstructed code rather than obtained from original designer records. Geometry alone often cannot identify a part’s true purpose or design rationale. Adding intent also changes text length, vocabulary, feature grouping, and description style.
  3. The evaluation is small and unstable. Each benchmark split contains only 200 parts, while Chamfer and IoU are averaged only over valid outputs. For the base model, some L2 metrics are computed from only 9–12 valid examples.
  4. Weak benchmarking and reproducibility. In Table 3, the performance gap between the three specialized text-to-CAD generation models (Text2CAD, Text-to-CadQuery, and CADFusion) and the method proposed in this paper is excessively large (with a CD value > 200 for the former). This substantial disparity may point to issues with the datasets or experimental setup, rendering it an unreliable indicator of the models' actual capabilities. Experiments should be conducted using more general-purpose datasets and consistent experimental settings.
Questions for authors
  1. Why does the proposed pipeline reconstruct all models in Build123d instead of preserving their native representations or using CadQuery?
  2. What is the specific difference between the “description normalization” and “design intent”?
  3. How does the paper demonstrate that the performance gains resulting from the description of "design intent" stem from the intent itself, rather than from additional textual information, feature grouping, or lexical variations?
  4. During program reconstruction, how are feature trees, parameter relations, and design intent inferred from final geometry?
  5. Why are the specialized baselines evaluated with CadQuery and 256³ voxelization, whereas IntentCAD is evaluated with Build123d and 128³ voxelization?
Reviewer_vjPKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents IntentCAD, a framework for text-to-CAD generation that studies what design information beyond geometry should be exposed in training data. Its central hypothesis is that design intent can act as privileged information. Its contributions are three-fold: a dataset of 446K text–program pairs from 223K parts, whose programs are normalized into features with named parameters and explicit relations and whose descriptions state each part’s use and design intent; controlled experiments showing the contribution of each data component; and a 4B model evaluated against both general-purpose and specialized text-to-CAD models. Although the proposed model achieves the best performance on some metrics, it lacks a comparison with the latest model (Table 3), e.g., GPT-5.6 Sol. In addition, it lacks further explanation for the inconsistent effect of the proposed data components (Table 2). In general, the reviewer finds the discussion and exploration regarding the use of design intent valuable, but it is limited to the training data and to single regular parts, which limits its contribution.

Critical strengths

The discussion and exploration regarding the use of design intent are valuable. By including such information in the training data, the model can produce higher-quality regular single-part CAD models. In addition, the contributed dataset could be beneficial to the community.

Critical weaknesses

In my opinion, the contribution of this work is limited. Although the finding that including design intent in the training prompt is insightful, the main contribution of this work lies mainly in the dataset. In addition, if the value of the dataset is to be verified, several base models could also be used for validation. If a larger-scale base model were used, would the gain brought by the data still be as significant? Furthermore, the exploration of the application of design intent is relatively limited. So far, it has only been verified that design intent is helpful during training. However, the paper does not verify or explore the potential benefits of design intent at other stages. For example, given a general-purpose model, would providing design intent at test time improve generation quality? If so, would this be a lower-cost solution?

Besides, regarding the experimental results, in Table 1, for different inputs, the effects of different data components (such as program normalization, description normalization, and design intent) on different evaluation metrics are inconsistent: some improve and some worsen. This does not well verify the role of the different data components, and the authors did not provide sufficient explanation.

Questions for authors
  1. Please add updated results for general-purpose models in Table 3, such as GPT-5.6 Sol and Claude Opus 5.
  2. If design intent were added to the test prompt of the best-performing general-purpose model in Table 3, what effect would it have?
  3. Are there specific categories or description standards for design intent and use?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.