- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies CAD program repair conditioned on a fixed Feature Plan. TopoRecover replays the faulty program with a deterministic state-machine interpreter, records completed profiles and their source-command intervals, and maps discrepancies between the constructed structure and the plan to local editing regions. A frozen LLM proposes constrained edits, followed by execution repair and a compiler fallback. Outputs must pass both plan checks and independent OpenCASCADE execution validation. On 347 repair inputs, the method achieves a 78.4% repair success rate, compared with 64.3% for the strongest adapted baseline.
Critical strengths
- A concrete and inspectable mechanism for localizing CAD repairs. The method links a fixed structural plan, saved completed profiles, and source-command intervals to identify construction blocks that execute successfully but violate the required structure. Within the supported error types, this provides direct, deterministic diagnostic evidence and enables preservation of unaffected construction content.
- Empirical support for both system effectiveness and the localization strategy. In addition to outperforming the evaluated repair methods, the paper fixes the patcher, validation procedure, fallback, and budget ceilings while replacing the region selector. Structural-discrepancy localization raises final repair success from 47.6% to 78.4%, supporting its effectiveness in this setting. Ablations of the LLM path and compiler fallback also demonstrate their complementary roles.
Critical weaknesses
Some claimed generalization contributions lack corresponding experimental reports. The contribution list and conclusion claim evaluation on held-out generators, unseen structure families, and additional feature operators, but I could not find the corresponding splits, protocols, or results in the main paper or appendix. The implemented scope is primarily sketches and extrusions, and the appendix explicitly describes the chamfer in Figure 1 as conceptual. These claims need supporting evidence or narrower wording that accurately reflects the demonstrated contribution.
Questions for authors
- Which data splits, experimental protocols, and results support the claimed evaluations on held-out generators and unseen structure families in the introduction and conclusion? If these experiments have not been conducted, should the corresponding contribution statements be revised?
- The contribution list and conclusion claim evaluation on additional feature operators. Please specify these operators, the error settings in the faulty programs, and the corresponding localization and repair results. In particular, for operations such as fillets and chamfers that act on solid edges or faces, please distinguish upstream sketch errors exposed by subsequent operations from errors in these features themselves: which cases can the current diagnosis based on profile structure and source intervals cover? Please align the generalization claims with the settings actually evaluated; if no additional-operator evaluation has been conducted, please narrow those claims accordingly.
2 comments
Updated Assessment
Thank you for the additional experiments and clarifications. The controlled results on unseen multi-profile constructions strengthen the evidence for the proposed localization mechanism. However, the cross-generator evaluation includes only four faulty programs, so evidence for broad transfer remains limited. The additional-operator experiments also establish effectiveness within controlled synthetic settings.
Despite these limitations, I find the core localization contribution and its empirical validation sufficient for acceptance. I will therefore maintain my rating of 3 (Weak Accept).
Thank you for recognizing the inspectable diagnosis and the controlled evidence for localization. We would like to address your questions individually as follows.
Q1: What experiments support the held-out-generator and unseen-structure-family claims, and should unsupported claims be revised?
A1: Our method can, in principle, transfer across generators and construction families because its diagnosis relies on command semantics and construction provenance rather than generator- or family-specific templates. We have now tested this transfer in supplementary experiments conducted during rebuttal, keeping the 7B repair policy frozen. We will revise the introduction and conclusion to align the generalization claims with these results and explicitly define the evaluated construction signatures.
Held-out generators. As a complement, we have evaluated point-cloud-conditioned CAD-Recode1.5 (Rukhovich et al., 2025; v1.5 checkpoint) using the frozen 7B repair policy and reference-derived structural plans. On 20 outputs, we preserved all 16 initially correct programs and repaired one of four incorrect programs, whereas execution-error-triggered repair recovered none.
Unseen construction families. We evaluated transfer on 27 controlled multi-profile repair cases whose input and target construction signatures were absent from the recorded exposure ledger. Each program contains 2–5 profiles. We kept the 7B patcher, fallback, validation and downstream budgets fixed, changing only the region selector. For the LLM-based selector, we use Qwen2.5-7B to choose regions from the program and plan.
By accurately localizing the faulty regions, we enable the same patcher to repair substantially more cases directly, demonstrating that our source-based diagnosis remains effective on these unseen multi-profile constructions without retraining.
Q2: What additional feature operators and error settings have been evaluated? Can the method distinguish upstream sketch errors from operator-local errors in fillets and chamfers?
A2: We would like to clarify that our localization mechanism can distinguish upstream sketch errors from errors within chamfer and fillet operations. Specifically, profile records identify upstream structural errors, while additional selected-edge and parameter records localize operator-specific errors to their corresponding command blocks. To validate this extension, we evaluated 80 controlled synthetic cases using a rule-based implementation, covering five scenarios (16 cases each):
All outputs passed constraint and CAD-kernel validation while preserving unaffected commands. These results demonstrate the effectiveness of the extension for both upstream and operator-local errors in the evaluated settings. We will revise the manuscript to explicitly report these supplementary experiments and align our generalization claims with the evaluated scope.