acceptodds
Withdrawn from ICLR 2027

Causal-CAD: Learning Causal Models for Root-Cause Analysis of Parametric CAD Errors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Parametric computer-aided design (CAD) stores both an ordered feature history and a graph of the references consumed by each feature. A local edit can propagate through this dependency graph and prevent distant downstream features from regenerating, while the CAD system often reports only where regeneration stopped. We present a system that learns executable causal models from Autodesk Fusion 360 dependency graphs before and after edits. Each model is a program defining a causal graph and structural functions over its variables. An LLM-guided quality-diversity search evolves these programs for several error families, and a live Fusion agent tests their explanations. Causal explanations from our system achieve a mean score of 0.238 on our repair task test bed, compared with 0.150 for an LLM-only explanation and 0.123 without an explanation. Performance also rises after validating the models in the live Fusion environment, demonstrating the value of integrating learning and acting in a real-world CAD setting.

Primary area:
causal reasoning
Keywords:
CAD, causality, evolutionary programming

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8XibOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how to identify the root cause of computer-aided design (CAD) errors and guide the corresponding repair. The authors first propose a system to read the dependency graph before and after edits and utilize an LLM to generate an executable causal model. Then, counterfactual perturbations are performed to give explanations and validate them in the Fusion 360.

Critical strengths
  1. The research on CAD error location and interpretation in this paper is quite meaningful and has practical value.
  2. This paper is easy to follow.
  3. The author validated the proposed method in real Fusion 360, enabling verification of the explanations provided by the method.
Critical weaknesses
  1. The novelty of this paper is limited. Causal models and counterfactual perturbations have already been extensively studied, and beyond these established components, the paper offers no other substantive methodological or theoretical innovations.
  2. The experimental evaluation is limited and does not yet provide sufficiently convincing evidence of the proposed method’s effectiveness. In particular, the small number of baselines makes it difficult to establish its advantages over existing approaches.
  3. In Equation 7, if the perturbation of variable r causes changes in other root variables as well, then the resulting counterfactual explanation might be inaccurate. This is particularly important in identifying the root cause of CAD errors. The author did not clarify whether the mutual influence among the root variables was isolated when generating the counterfactual explanation.
  4. The description of the executable causal model is not clear enough. The differences between the model mentioned in this paper and the general structural causal model have not been clearly explained. In the "Executable causal models", the role of the intermediate variable I has not been demonstrated.
  5. The experimental setup is not clearly described. The criteria for selecting CAD error cases and the number of cases are not explicitly stated. Moreover, the existing experiments do not demonstrate the effectiveness of the verification feedback.
Questions for authors
  1. As the number of learning-acting cycles increases, both the search computation and validation feedback of the proposed method also increase. How to determine whether the performance improvement comes from the feedback mechanism?
  2. Could the authors briefly introduce the main types of CAD errors? This would help readers without domain expertise better understand the proposed approach.
Reviewer_MQAvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Causal-CAD, a framework for root-cause analysis of parametric CAD errors using executable causal models learned from Fusion 360 feature dependency graphs. The method employs LLM-guided evolutionary program search to construct causal models that predict CAD failures, and uses counterfactual analysis to identify likely root causes and generate explanations. A learning–acting loop further refines these models using feedback from an LLM agent interacting with the live Fusion 360 environment. The authors evaluate the framework across eight CAD error families using downstream repair performance. Causal explanations achieve a mean repair score of 0.238, compared with 0.150 for LLM-only explanations and 0.123 without explanations. Additional experiments show improved repair performance over successive learning–acting checkpoints and explore shared semantic patterns across learned causal models.

Critical strengths
  1. Meaningful problem and structured approach. The paper addresses root-cause analysis of parametric CAD errors, where failures can propagate through feature dependencies and the reported error may not reveal the underlying cause. Representing these dependencies through executable causal models provides an interpretable approach to diagnosing complex CAD failures.
  2. Practical validation in a real CAD environment. Rather than evaluating explanations solely through prediction accuracy, the authors assess their usefulness through downstream repair tasks using the Fusion 360 API. The causal explanations improve the mean repair score from 0.150 with LLM-only explanations to 0.238. This provides encouraging evidence that structured diagnostic information can improve the effectiveness of CAD repair agents.
Critical weaknesses
  1. Insufficient validation of causal reasoning and key components. The causal models are primarily optimized for error prediction using Macro-F1, which does not necessarily establish that the learned relationships represent true root causes. Furthermore, the learning–acting experiment combines additional evolutionary search with live Fusion feedback, without an ablation that isolates the contribution of the acting stage. This makes it difficult to determine whether the improvements come from reliable causal reasoning or from additional search and diagnostic information.
  2. Limitations of the downstream repair evaluation. The repair metric relies partly on structural statistics such as body counts, feature counts, and timeline length, rather than directly measuring geometric correctness or preservation of design intent. Consequently, a repair could receive a favorable score while introducing unintended geometric changes. Without stronger geometry-based evaluation or more systematic expert validation, the reported improvements provide limited evidence of reliable CAD repair.
Questions for authors
  1. Causal validity and ablation. Since the causal models are optimized primarily for error prediction, what evidence supports the correctness of the identified root causes rather than correlations with the observed failures? Could the authors also clarify how much of the learning–acting improvement comes specifically from live Fusion feedback, compared with additional evolutionary search under the same budget?
  2. Repair evaluation. The overall repair score relies partly on structural statistics such as feature counts and timeline length, which may not capture unintended geometric changes. Could the authors provide additional evidence, such as geometry-based comparisons or systematic expert assessments, showing that the higher repair scores correspond to genuinely correct CAD repairs that preserve the intended design?
Reviewer_WPdKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Causal-CAD learns executable diagnostic programs from Fusion dependency graphs before and after an edit. Root variables extract graph measurements; intermediate functions and an error function form an acyclic computation graph. LLM-guided quality-diversity search selects programs by predictive macro-F1. Counterfactual root perturbations rank explanatory factors, and a live Fusion agent investigates hypotheses and feeds mechanism and vocabulary updates into further search. Eight error families are evaluated through explanation-guided repair. The reported mean composite repair score is 0.238, versus 0.150 with LLM-only explanations and 0.123 without explanations. Later learning–acting checkpoints improve this score. The important scientific question is whether the learned programs and their interventions capture repair-relevant CAD mechanisms, beyond predicting failure symptoms.

Critical strengths
  1. The task and representation address a real weakness of parametric CAD diagnostics. Feature-history failures can emerge downstream of the edit that invalidated a reference or operation precondition. Explicit programs over before/after dependency information are more inspectable than an unconstrained narrative: intermediate quantities and parent accesses can be checked against a concrete design. Retaining diverse candidate explanations is also sensible when several mechanisms predict the same error family.
  2. Live investigation provides a meaningful corrective channel beyond offline predictive fitness. The acting agent can inspect geometry and references, perform a disposable edit, and reject every candidate hypothesis. The repair comparison holds the downstream model and tool interface fixed and shows a useful average gain. The authors also report expert counterexamples and explicitly acknowledge that predictive fitness may reward correlates. These qualifications are scientifically valuable, although they do not settle causal validity.
Critical weaknesses
  1. The stated causal score can assign influence to an irrelevant root, and sampled worlds may be unrealizable CAD states. Section 4 perturbs all roots and defines s(r) = P(prediction changes | r changes). Consider Y = X with an irrelevant root Z, observed X = Z = 1, and independent fair-bit resampling. Conditional on Z changing, X changes in half the worlds, so Equation (7) gives s(Z) = 1/2 although Z has no effect on Y. Joint sampling itself is not the problem: the cited CESM permits multiple variables to vary, but uses an effect-size statistic rather than this conditional outcome-flip fraction (Quillien and Lucas, DOI: 10.1037/rev0000428). If the implementation uses paired interventions or a different statistic, the paper must provide it. More fundamentally, roots are overlapping measurements of the same graph, not demonstrated independent intervention variables. Removed-reference totals and subtype counts cannot generally be changed independently while remaining consistent with one CAD edit. The paper recomputes downstream program variables but does not explain how root assignments preserve graph invariants or correspond to kernel-executable edits. A counterfactual that flips the diagnostic program can therefore lack a realizable repair interpretation. Sampling more worlds cannot correct this semantic problem.
  2. Prediction-selected computation graphs are not yet identified CAD failure mechanisms. Equations (3)–(6) enforce executable parent order and optimize macro-F1; neither identifies causal direction among graph-derived measurements. For example, downstream feature orphaning can predict an error while being a consequence of the upstream reference loss that should be repaired. A variable declared as a computational root need not be a causal root in the CAD history. Acyclic source-code dependencies and semantic variable names do not close that gap. The live acting loop can reduce it, but its verdicts are incorporated through LLM feedback rather than an explicit objective matching predicted and observed intervention outcomes. The repair score does not resolve mechanism validity. Equations (10) and (13) combine error clearance with similarity of body, sketch, feature, and component counts plus a removal penalty. Counts can agree while geometry, selected references, or intended edits differ substantially. The metric is a useful proxy, but s_root does not directly measure root-cause localization, and Appendix A does not specify a quantitative check that the triggering edit is retained. The contribution is presently better established as learned diagnostic assistance than as verified causal root-cause analysis.
Questions for authors

The following are questions about the validity of the current method.

  1. Does the implemented estimator literally follow Equation (7)? What score does it assign to the irrelevant-root example above? Please provide the exact estimator and distinguish it from the CESM effect size.
  2. Which roots are independently intervenable, what constraints preserve consistency among overlapping graph measurements, and how is a counterfactual assignment converted into a realizable Fusion edit?
  3. Can the predicted effects of selected interventions be compared with observed regeneration outcomes, including null interventions and downstream symptom variables? This directly tests the causal claim rather than explanation fluency.
  4. How are preservation of the triggering edit and intended geometry checked independently of feature counts? Please separate actual repair success, structural similarity, and root-cause localization, and clarify how much of the gain remains without causal ranking.

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

Related papers

Discussion (0)

Sign in to comment.