acceptodds
Withdrawn from ICLR 2027

AIM-CAD: Causality-Inspired Auditing of Historical Edits for Training-Free CAD Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Language-guided parametric CAD editing requires realizing user-specified modifications while preserving unaffected content and maintaining a valid, editable model. However, obtaining large-scale, high-quality editing supervision is costly, while task-specific fine-tuning requires additional model-specific adaptation. This motivates training-free editing that reuses historical edits as external experience. Such reuse remains challenging: new requests may differ from available examples in instruction or CAD structure, and historical traces may contain redundant or context-specific operations. Moreover, plausible edit plans may not translate faithfully into executable CAD operations. We propose Audited Intent–Operation Memory for CAD Editing (AIM-CAD), an experience-guided, training-free framework for reusing historical edits. AIM-CAD builds an audited memory by organizing historical edits as intent–operation graphs and applying causality-inspired interventions to check instruction-level executable sufficiency and prune redundant operation attributions. These interventions assess functional support within a fixed CAD program. Relevant audited experience is then retrieved across instruction semantics and CAD structure. A Graph-Guided Dual-Output Generator produces both structural and executable realizations, whose consistency is checked by Graph–Script Consistency Validation before output. Experiments on adapted CAD-Editor and BenchCAD benchmarks across multiple LLM backbones demonstrate the effectiveness of AIM-CAD. Code will be available at https://anonymous.4open.science/r/aim-cad-4F2B/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Aided Design, Text-based Editing, CAD Editing, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jpzuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AIM-CAD proposes a training-free framework for LLM-based parametric CAD editing. It decomposes natural-language editing instructions into atomic intents and constructs intent–operation graphs from historical CAD edits. Through causality-inspired sufficiency and necessity checks, the framework audits historical editing experiences to identify functionally supported intent–operation relationships and reduce redundant attributions. These audited experiences are subsequently retrieved to guide a frozen LLM in generating both structural edit graphs and executable CAD scripts, with cross-representation consistency validation improving editing reliability.

Critical strengths

The paper explores a training-free approach to improving the ability of LLMs to perform parametric CAD editing. By reusing historical editing experiences, the proposed framework avoids the need for task-specific fine-tuning. This is a practical direction with potential applications in industrial CAD workflows.

Critical weaknesses
  1. The semantic correctness of the editing results is not sufficiently verified. During historical memory construction, the proposed auditing mechanism checks whether the editing operations can reproduce the target CAD model, but it does not necessarily guarantee that these operations accurately reflect the corresponding natural-language instructions. A similar issue exists during inference. The Graph–Script Consistency Validation ensures that the two generated outputs are executable and consistent, but both could still implement the same incorrect modification. Without an explicit mechanism to verify whether the final result actually follows the user's instruction, the reliability of the framework remains a concern.

  2. The CAD editing tasks appear relatively simple. The experiments mainly focus on editing simplified, JSON-based representations of parametric CAD models. While this provides a reasonable setting for evaluating the proposed method, it remains unclear how well the framework would handle more complex CAD models or real-world editing workflows in commercial CAD software. This limits the evidence supporting its practical applicability.

Questions for authors

How does AIM-CAD ensure that the final edited CAD model faithfully follows the user's natural-language instruction? While Graph–Script Consistency Validation verifies executable validity and cross-representation agreement, both outputs are generated by the same LLM and may consistently implement an incorrect modification. What mechanism prevents such semantically incorrect but structurally valid edits from being accepted?

Reviewer_RUrcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AIM-CAD, a training-free framework for CAD editing. Instead of fine-tuning, AIM-CAD reuses historical edit examples through an external memory, so at inference time, the system first retrieves a relevant audited example and then generate an edit graph and a Python script. Experiments are conducted on an adapted CAD-Editor benchmark and a BenchCAD CodeEdit subset

Critical strengths
  1. The setting is interesting and practical since paired language/CAD editing supervision is expensive
  2. The paper shows strong empirical gains in execution reliability according to the experimetns
  3. The concretes examples help to understand the paper.
Critical weaknesses
  1. First of all, the paper content seems to be heavily AI generated. Lots of the text read AI heavy and I also highly doubt the figures are ai generated.

  2. The empirical gains may come from the overall retrieval + execution-validation pipeline rather than the proposed causal component, especially according to the ablation study.

  3. The main results, table 1, didn't compare with any baselines, but only vanilla and sequence editing. This weakens the argument of the paper.

  4. Most evaluaiton metrics are only computed on each method's valid predictions, meaning the subsets differ across methods.

  5. The proposed methods involves a very complex pipeline including graph construction, retrieval, joint graph/script generation, etc, so very clearly it is much more expensive than vanilla prompting. The cost is not analyzed and justified.

Questions for authors

How much of AIM-CAD's improvement comes specifically from causal/intervention auditing?

Can you report runtime and cost?

Reviewer_E86nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AIM-CAD, an experience-guided training-free framework for language-guided parametric CAD editing. The key motivation is that high-quality CAD editing examples are (1) high-quality CAD editing examples are costly to obtain and historical edits cover limited use cases, (2) the intent-operation evidence in historical edits are ambiguous and noisy, and (3) the gap between edit plans and executable realizations are hard to measure. To address these challenge, AIM-CAD propose to store these historical editing experience through historical editing graph, and retrieve relevant experience by jointly considering the instruction and CAD structure. Specifically, it decomposes editing instructions into atomic intents, aligns these intents with JSON patches extracted from source–target CAD pairs, and performs causality-inspired auditing through executable interventions. During inference, AIM-CAD retrieves relevant historical examples and a frozen LLM then jointly generates an edited graph and an executable Python script. Their consistency is validated before output. The framework is evaluated on adapted CAD-Editor and BenchCAD benchmarks across multiple LLM backbones, demonstrating improvements in output validity, structural similarity, and unintended change rate.

Critical strengths

The paper addresses a practically meaningful problem and presents a reasonable engineering intuition for reusing historical editing experiences without task-specific model training.

The paper is generally well-written and easy to follow.

Critical weaknesses

Limited validation of intent–operation alignment. The introduction motivates the intervention-based causal auditing as a mechanism that could identify which operations support individual editing intents (L075, L083). However, the sufficiency check only verifies whether the patches associated with an instruction collectively reproduce the historical target, while the necessity check identifies redundant patches based on executable equivalence. Neither directly verifies whether individual intents are correctly aligned with their operations. It remains unclear whether the observed improvements come from more accurate attribution or simply filtering unreliable historical examples.

Limitations of graph–script consistency validation. Although the consistency between the historical edit graph and the Python script is validated through execution, both outputs may consistently implement the same incorrect edit. Furthermore, the script-only ablation removes both graph generation and consistency validation, making their individual contributions unclear.

Insufficient evaluation of editing success. UCR does not penalize missing modifications and is defined as zero when no edit is performed. Therefore, a low UCR does not necessarily indicate successful editing. Moreover, most quality metrics are computed only over valid predictions, which can differ substantially across methods. Additional edit-completion metrics and comparisons on matched evaluation subsets would strengthen the conclusions.

Missing computational cost Analysis. Although AIM-CAD and the baselines share the same validation budget, AIM-CAD may spend higher inference costs due to reference-conditioned prompting, dual-output generation, and additional validation. The offline cost of constructing and auditing historical memory is also not reported. A cost analysis would help assess whether the performance gains justify the additional complexity.

Questions for authors

See weakness.

  1. For BenchCAD, the original benchmark [1] explicitly uses normalized accuracy, based on the improvement in voxel IoU over the original CAD, as its primary metric for the CodeEdit task. Could the authors clarify why this metric is not included in the evaluation and report it on the adapted benchmark?
  2. How frequently does graph–script consistency validation reject generated outputs? Could the authors provide a breakdown of the failure types detected by this mechanism?

[1] Zhang, Haozhe, Kaichen Liu, Miaomiao Chen, Lei Li, Shaojie Yang, Cheng Peng, and Hanjie Chen. "BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD." arXiv preprint arXiv:2605.10865 (2026).

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

Related papers

Discussion (0)

Sign in to comment.