acceptodds
Withdrawn from ICLR 2027

RA-CAD: Learning Post-Execution Critique for State-Aware Text-to-CAD Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-to-CAD generation translates natural-language design intent into editable and executable parametric computer-aided design (CAD) codes, reducing the expertise and effort required for manual modeling. Existing methods incorporate fixed, externally supplied, prompt-induced, or separately optimized critique mechanisms to optimize the generation process, but they do not necessarily optimize how feedback is interpreted and translated into effective corrective actions throughout the generation process. To bridge this feedback-utilization gap, we present RA-CAD (ReAct Agent for CAD), a state-aware agent that interacts with the CAD environment through a Generate–Execute–Critique–Rewrite loop. At each iteration, RA-CAD executes the current code and observes its outcome. Conditioned on the design instruction, current code, and execution feedback, the agent then generates an explicit post-execution critique as an intermediate policy action. This critique either validates the current result for termination or provides revision-oriented guidance that conditions the next rewrite. CAD Code Bootstrapping (CCB) first establishes fundamental parametric CAD coding capabilities through supervised fine-tuning. Feedback-Driven Agent Optimization (FAO) subsequently applies trajectory-level Group Relative Policy Optimization to both policy-generated code and critique sequences, assigning terminal F1 and Chamfer Distance rewards to the complete interaction trajectory. This formulation makes critique an outcome-aligned, learnable policy decision rather than an unoptimized auxiliary output. Experiments on CADFusion and Text2CAD show that RA-CAD achieves state-of-the-art execution validity and geometric quality compared with existing methods and strong proprietary language models, demonstrating the effectiveness of the proposed state-aware text-to-CAD agent.

Primary area:
reinforcement learning
Keywords:
Text-to-CAD, ReAct Agent, Feedback-Driven Optimization

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_wZmUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the CAD generation task and points out that existing heuristic feedback mechanisms do not necessarily translate feedback into effective corrective actions to rectify generation errors. To address this issue, the paper proposes RA-CAD, which organizes CAD generation into a Generate–Execute–Critique–Rewrite loop. Specifically, it generates explicit post-execution critiques to guide subsequent CAD code rewriting.

Critical strengths
  1. The idea of introducing explicit critiques to improve CAD code rewriting is interesting.
  2. Figure 3 clearly illustrates the workflow of the Generate–Execute–Critique–Rewrite loop.
Critical weaknesses
  1. Introducing the Critique module is a core design of this paper. However, the implementation details of the Critique module appear to be missing. Section 4.1.2 mentions Meta-Llama-3-8B-Instruct as the base model, but it is unclear whether this refers to the Generation module, the Critique module, or both. Moreover, is the entire Generate–Execute–Critique–Rewrite loop jointly trained using the reward described in Section 3.3.3? Is the Critique module trained separately? Furthermore, I would like to better understand the content of the critiques and how they are designed. Although I noticed that the Appendix provides the “Prompt of the Critique Module,” I would still appreciate a discussion of the motivation behind the critique design, along with concrete examples of the critiques generated by the module.

  2. The experimental results appear strange, and I would like the authors to provide explanations.

    • The results in Table 1 are strange. In the original papers on Text2CAD [1] and CADFusion [2], the reported invalidity ratios (IR) are very low (below 10%). However, in Table 1 of this paper, the IR are considerably higher: above 60% for Text2CAD and above 20% for CADFusion. This discrepancy is quite strange and needs explanations.

    • In related RL-based CAD generation studies[2][3], the supervised fine-tuning (SFT) stage typically leads to substantial improvements in the invalidity ratio (IR). Similarly, as mentioned in Sections 3.2 and 3.3 of this paper, SFT is intended to teach the model CAD syntax and establish its basic capability to generate executable CAD code. However, in Table 3, the w/o FAO setting, which represents the results obtained using SFT alone, yields an extremely high IR of 90.01%. This suggests that the model still struggles to generate executable CAD code even after SFT, which is counterintuitive. I would appreciate a detailed explanation of this result.

  3. Some more advanced works[4-6] on CAD generation should be included and discussed in the Related Work.

[1] Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts.NIPS 2024

[2] Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models.ICML 2025

[3] cadrille: Multi-modal cad reconstruction with online reinforcement learning. Arxiv 2025

[4] CADMorph: Geometry-Driven Parametric CAD Editing via a Plan-Generate-Verify Loop. NIPS 2025

[5] CAD-Tokenizer: Towards Text-based CAD Prototyping via Modality-Specific Tokenization. ICLR 2026

[6] Plan-then-Act: Bi-level CAD Command Sequence Generation. ICLR 2026

Questions for authors
  1. Please clarify the implementation details of the Critique module.
  2. Please provide an explanation for the strange experimental results.
Reviewer_uzS6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies text-to-CAD generation, targeting executable and geometrically accurate CAD programs. RA-CAD introduces learnable post-execution critique to guide iterative code revision and termination. Its CCB–FAO pipeline combines SFT with trajectory-level GRPO to jointly optimize generation, critique, and rewriting. Baseline comparisons and ablations on CADFusion and Text2CAD show improvements in execution validity and geometric quality.

Critical strengths
  1. The paper addresses a practically relevant problem, with a clear motivation for feedback-guided CAD refinement.
  2. Experiments on two datasets show improvements in execution validity and geometric quality.
Critical weaknesses

Overall, I think the paper’s contributions appear largely incremental, primarily involving the integration of existing techniques and their application to the text-to-CAD task.

  1. Limited methodological novelty of learnable critique. Learnable critique and feedback-driven self-correction have clear precedents in agent research [1, 2], and existing CAD methods already incorporate learned iterative refinement [3, 4]. The main distinction of this work is to separate post-execution critique into an explicit decision stage that guides CAD code revision and termination. However, its substantive advantage over existing learned refinement strategies has not been sufficiently established.

  2. Training largely follows established paradigms. CCB and FAO combine standard supervised fine-tuning, trajectory-level GRPO, and CAD-specific quality rewards. Their independent methodological contribution appears limited, and the paper should more clearly explain what additional value this combination provides over commonly used SFT&RL training pipelines.

Reference:

[1] Xie et al. Teaching Language Models to Critique via Reinforcement Learning. ICML, 2025.

[2] Gehring et al. RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning. ICML, 2025.

[3] Gong et al. TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning. Findings of ACL, 2026.

[4] Hu et al. IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing. EMNLP, 2026.

Questions for authors
  1. What evidence shows that the explicit, learnable critique stage improves over existing critique strategies?

  2. What is the specific methodological contribution of CCB–FAO beyond standard SFT and trajectory-level GRPO?

Reviewer_3LPYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents RA-CAD, a text-to-CAD agent that uses execution feedback to critique and revise CAD code. Following supervised fine-tuning, GRPO jointly optimizes code generation and critique. Experiments on CADFusion and Text2CAD show improved execution validity and geometric quality.

Critical strengths
  1. The paper presents a well-motivated framework in which the model learns to generate CAD code, review execution results, and revise the code when needed.
  2. Experiments on CADFusion and Text2CAD show clear improvements in execution validity and geometric accuracy over the evaluated baselines.
Critical weaknesses
  1. Limited novelty. CADDesigner, ToolCAD, and IterCAD already incorporate feedback-driven refinement. The proposed approach’s novelty in feedback use and policy learning remains unclear.
  2. Limited evaluation of complex modeling. The evaluation should cover a broader range of modeling operations, including revolves, chamfers, and shelling.
  3. Limited baseline comparisons. The evaluation should include comparisons with recent CAD methods (e.g., CAD-Coder, ReCAD, ToolCAD, and IterCAD) and newer general-purpose LLMs (e.g., GPT-5.6 or GPT-6).
  4. Insufficient experimental settings. The method uses iterative execution and revision, whereas LLM baselines use single-pass generation, limiting comparability. The ablation jointly removes Critique and iterative revision, obscuring their individual contributions.
  5. Insufficient evaluation and analysis. The treatment of non-executable outputs in Avg CD requires clarification, as differing invalidity ratios may affect comparability. Evaluation should cover unseen structural combinations and longer sequences. The CD and IR comparisons between fixed G&R and fixed Crit in Sec. 4.4 are reversed relative to Table 3.
  6. Limited validation of critique and rewriting reliability. Appendix F presents successful corrections but does not assess false acceptance or revision-induced errors.
Questions for authors
  1. Clarify the differences from prior methods. Clarify how the proposed method differs from CADDesigner, ToolCAD, IterCAD in the use of execution feedback, critique training, and revision decisions. Justify the necessity of the proposed training strategy.
  2. Add comparisons with recent methods. Compare against recent reinforcement learning and agent-based CAD methods, including the iterative correction approaches discussed above, as well as strong general-purpose models.
  3. Add results on complex models. Please evaluate on at least one complex CAD dataset. Report results for revolves, chamfers, shelling, and combinations of operations.

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

Related papers

Discussion (0)

Sign in to comment.