acceptodds
Under review as a conference paper at ICLR 2027

PhyCAD: Physics-Aware Text-to-CAD Generation via Hierarchical Reinforcement Learning

Abstract

Language-driven computer-aided design (CAD) generation enables the creation of editable parametric models from text, yet geometric correctness alone does not ensure physical feasibility. We present PhyCAD, a framework for physics-aware Text-to-CAD generation that trains language models to produce executable CadQuery programs under geometric and physical constraints. To support this setting, we construct a Text-to-CAD dataset that links parametric CAD models with geometric specifications, loading conditions, and physical constraints. We further develop an automated CAD–CAE verifier that combines program execution, geometric checks, and finite element analysis to provide feedback for model training. A key learning challenge is to prioritize joint satisfaction of geometric and physical constraints while retaining informative feedback on partial progress. A flat aggregation of heterogeneous rewards can obscure this priority, allowing improvements in individual criteria to compensate for unmet design requirements. We therefore introduce a Hierarchical Balanced Advantage for Group Relative Policy Optimization, grouping candidates into levels according to their verification status and combining between-level progress with objective-wise peer ranking within each level. This formulation preserves cross-level priority in the training signal while providing fine-grained feedback for satisfying constraints and improving feasible designs. Experiments across multiple language models show that PhyCAD consistently improves joint geometric and physical constraint satisfaction. These results demonstrate the value of verifiable engineering feedback for advancing language-driven CAD generation beyond geometric reconstruction toward physics-aware design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Conditioned CAD Generation, Text-to-CAD

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_WZW5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhyCAD, a framework for physics-aware text-to-CAD generation that considers both geometric and physical constraints. Its contributions are three-fold: a text-to-CAD dataset, an automated CAD-CAE verifier, and a novel reward mechanism called Hierarchical Balanced Advantage for GRPO. Specifically, the CAD-CAE verifier combines three-step checking, including program execution, geometric checks, and finite element analysis. The Hierarchical Balanced Advantage, which is the most important contribution of this work, is designed to ensure that the output satisfies both geometric and physical constraints. By combining a between-level advantage with a within-level comparison, it effectively mitigates the drawback of flat aggregation of heterogeneous rewards—namely, allowing improvements in individual criteria to compensate for unmet requirements. The authors conducted extensive experiments on both closed-source and open-source LLMs. The results demonstrate the effectiveness of the proposed PhyCAD. The ablation results shown in Table 2 also verify the performance of the hierarchical GRPO.

Critical strengths

The discussion of the drawback of flat reward aggregation in GRPO is insightful. The authors correspondingly propose a solution, namely the hierarchical GRPO. Accompanied by feedback from the CAD-CAE verifier, it ensures that the output satisfies both the geometric and physical constraints.

Critical weaknesses

None

Questions for authors

In Table 2, the authors report results for Geometry GRPO, Physics GRPO, and Hierarchical GRPO. However, they do not provide implementation details for these variants. For example, how many steps each RL stage involves and what the reward formulation is for each stage. In addition, does the training order affect the final results, e.g., physics GRPO followed by geometry GRPO versus the reverse?

2 comments
AuthorsOpenReview ↗

Reward Formulation and Training Strategy Analysis

We sincerely thank the reviewer for the highly positive evaluation and strong recognition of our contributions. We also appreciate your constructive questions and the opportunity to further clarify our method.

(1) Reward formulation. The stage-specific rewards are illustrated in Fig. 3. For staged training, we use the within-level PeerRank term of Eq. (5):

Aistage=λwA~i=λw∑kwkrik.A_i^{\mathrm{stage}}=\lambda_w\widetilde{A}_i=\lambda_w\sum_k w_k r_{ik}.

Specifically:

  • Geometry GRPO (Fig. 3, L1): Uses bounding-box closeness and geometric-semantic recall.
  • Physics GRPO (Fig. 3, L2–L3): Uses stress, displacement, and mass closeness at L2, and their corresponding feasibility margins and mass efficiency at L3.
  • Hierarchical GRPO (Fig. 3, L0–L3): Uses the complete Eq. (5), combining between-level advantages with within-level PeerRank across all four feasibility levels.

We will further formalize these stage-specific reward formulations with explicit mathematical expressions in the supplementary material.

(2) Training configurations. Table 2 compares two training strategies, both initialized from the same SFT model:

StrategyTraining sequenceSteps
DirectHierarchical GRPO7,000
StagedGeometry → Physics → Hierarchical GRPO14,000 → 14,000 → 7,000

All Hierarchical GRPO stages use 7,000 steps. We also extended training with the individual Geometry and Physics objectives to 14,000 steps, observing no substantial further improvement. This suggests that longer optimization with isolated objectives is insufficient to replace the joint hierarchical feedback. This further validates the effectiveness of our hierarchical reward design over isolated geometric or physical optimization.

(3) Effect of training order. We additionally evaluate the reverse training order (Physics → Geometry → Hierarchical GRPO) under identical training budgets and evaluation settings in Qwen2.5-3B.

Training OrderStageValid CAD ↑BBox ↑CAE ↑Physics ↑Joint ↑Code Tokens ↓
DirectHierarchical GRPO98.5588.5096.3240.9439.49195.46
Geometry → PhysicsGeometry GRPO95.2972.3791.5525.8822.86191.60
Geometry → Physics GRPO94.7275.2791.4030.3327.63188.03
Geometry → Physics → Hierarchical GRPO99.2392.0597.4543.3442.01184.85
Physics → GeometryPhysics GRPO93.7172.8790.3130.4827.28190.67
Physics → Geometry GRPO94.9674.5991.4029.4726.86188.06
Physics → Geometry → Hierarchical GRPO99.0589.7496.6842.4340.82186.00

Both training orders achieve substantial improvements after Hierarchical GRPO, with only modest differences in final joint feasibility (42.01% vs. 40.82%).

AuthorsOpenReview ↗

Follow-up on Rebuttal Clarifications and Additional Experiments

Dear Reviewer WZW5,

Thank you for your positive assessment and for recognizing the value of the hierarchical feedback formulation in addressing the limitations of flat reward aggregation. We appreciate your support for our work.

In response to your questions, we have clarified the objectives used by Geometry GRPO, Physics GRPO, and Hierarchical GRPO, specified their training schedules, and added the reverse-order experiment. Both staged training orders improve after the final Hierarchical GRPO stage, reaching joint feasibility of 42.01% and 40.82%, respectively. We would appreciate your feedback on whether these clarifications and additional results address your remaining questions.

The ongoing discussion also concerns whether PhyCAD makes a meaningful contribution beyond applying the established SFT+GRPO training paradigm. Your review identified the organization of heterogeneous CAD–CAE feedback as an important contribution and recognized the supporting ablation evidence. We would particularly value your assessment of this methodological contribution in light of the additional explanations and experiments.

If you find our responses satisfactory, we would be grateful if you could reaffirm your support for acceptance and reflect your updated assessment in your review and the final discussion. Your perspective on the contribution and the extent to which your concerns have been addressed would be especially helpful.

Thank you again for your time and consideration!

Best regards,

The Authors of Submission 33572

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

PhyCAD trains language models to generate CadQuery programs under geometric, displacement, stress, and mass constraints. An automated verifier assigns four levels, from invalid construction to joint feasibility. GRPO combines a group-centered level term with bounded objective-wise ranks among candidates at the same level. The final policy generates a design directly, without an iterative simulation loop at inference. The dataset derives inverse-design tasks from 77,067 reference designs under a connected-solid, linear-elastic steel setting. Evaluation on 3,373 held-out tasks reports improved joint feasibility across five small models; Ministral-3-3B increases from 37.36% after SFT to 58.05%. The method should be assessed as feasibility-prioritized learning with additional margin preferences, rather than as demonstrated constrained mass minimization.

Critical strengths
  1. The method addresses requirement satisfaction rather than only geometric imitation. Withholding exact reference dimensions while providing admissible ranges and physical limits creates a meaningful family of possible designs. Using CAD execution and CAE outcomes as feedback is a useful route toward learning to produce feasible parametric programs directly. The improvement across all five models supports the complete pipeline, and avoiding repeated simulation during generation is a practical benefit.
  2. The feasibility ordering is explicit and mathematically inspectable. The bound A_i - A_j >= lambda_L - 2 lambda_w follows from bounded within-level ranks. The chosen values 1.0 and 0.25 ensure a higher sampled advantage for a higher verification level. Capping satisfied constraints at the infeasible level also avoids rewarding excess performance on an already satisfied quantity while another requirement fails. The paper correctly avoids interpreting the bound as a guarantee that generated designs are feasible.
Critical weaknesses
  1. The feasible-level advantage does not implement the stated mass objective, and its algorithmic distinction needs sharper definition. Equation (1) minimizes mass within the feasible set. Section 3.3 instead averages ranks for mass efficiency, stress margin, and displacement margin at level 3. Consider two feasible candidates: A is lighter but has smaller, uncapped margins; B is heavier with larger margins. Their objective ranks can be (+1, -1, -1) and (-1, +1, +1), so PeerRank favors B and can give A negative advantage. Capping margins at 10% does not remove this counterexample before saturation. The authors acknowledge the combined preferences, but this does not reconcile them with the mass-only formulation. For zero-mean centered ranks within each level, define R_i = lambda_L L_i + lambda_w PeerRank_i. Equation (5) is then exactly R_i minus its group mean. This identity does not invalidate the method, but shows that its ordering result follows from bounded reward shaping. The technical contribution therefore needs to be located in the CAD-specific feedback and rank construction, relative to simpler bounded shaping and feasibility-first or lexicographic alternatives. With group size 4, singleton levels additionally receive no within-level signal. Table 2 does not isolate PeerRank from the level term or specify the flat reward sufficiently.
  2. The verifier does not yet define how engineering intent remains fixed as the geometry changes. Section 3.2 requires prescribed geometric constructions, but Appendix A mainly describes extracting dimension-free semantics from the reference program. Presence of a hole or shell operation is not sufficient to establish its location, interface role, or resulting topology in a new solid. The issue is especially consequential because local hard dimensions are largely absent by design. A shape can satisfy a global box and broad feature vocabulary while changing the intended load path or functional interface. A geometric acceptance predicate on the executed solid is needed to justify the feasible set in Equation (1). Appendix A derives fixtures and loading from terminal surface groups in the reference. How these regions map to a changed design is unclear. Reselecting terminals can change the physical problem; transferring them requires a correspondence rule. The 99th-percentile stress is a legitimate defined statistic, but certifies that statistic rather than peak-stress compliance. Together with unspecified mesh verification, these points limit interpretation of the verifier as engineering feasibility. The CAE rate exceeding BBox in Table 1 also requires an explicit distinction between training gates and evaluation denominators.
Questions for authors

These requests concern the stated optimization and verification mechanisms.

  1. Is the intended feasible-design objective mass minimization or a trade-off with safety margins? Please reconcile Equation (1) with the two-candidate counterexample, or modify the feasible-level advantage to respect the stated objective.
  2. What is distinctive beyond group-centered bounded reward shaping? Please define the rank convention and compare level-only feedback with PeerRank under the same verifier. A relevant conceptual reference is Skalse et al., Lexicographic Multi-Objective Reinforcement Learning, arXiv:2212.13769.
  3. How are required topology, feature counts, locations, and relations checked in the executed solid? How are the same fixture and loading semantics maintained when a candidate changes geometry or topology?
  4. Can a representative sample of accepted designs be independently rechecked for semantic compliance and refined-mesh responses, with mass reported among feasible designs? Please also explain the CAE/BBox denominator discrepancy.
5 comments
AuthorsOpenReview ↗

Design Objective and Hierarchical Reward Design

We sincerely thank the reviewer for the thoughtful and constructive feedback. We greatly appreciate these insights and address the concerns below.

1. Feasible-design objective.

Our intended objective is to minimize mass without driving designs excessively close to their physical limits. The mass constraint prevents excessive material usage as a trivial solution to stress and displacement requirements, as illustrated by the DeepSeek example in Fig. 4. Meanwhile, the safety-margin terms in Level 3 serve as soft constraints that discourage designs operating too close to their stress and displacement limits, rather than imposing additional hard feasibility requirements. We acknowledge that Eq. (1), which emphasizes mass minimization under physical constraints, does not sufficiently convey this design consideration. We will clarify the formulation by explicitly distinguishing the mass-minimization objective from the safety-margin preferences introduced through the reward, without changing the underlying optimization method.

2. Distinctiveness beyond bounded reward shaping.

We agree that Eq. (5) can be written in the group-centered form identified by the reviewer. In CAD–CAE reinforcement learning, the challenge is to construct informative rewards from the feedback available at each verification stage. Figure 3 specifies which verifier outputs are used at each level and how they are scored. HBA turns the verification sequence into a reward hierarchy, giving priority to candidates that pass more stages and using these scores to rank candidates objective by objective within each level.

Level-only ablation. To isolate the contribution of PeerRank, we remove the second (within-level PeerRank) term from Eq. (5), retaining only the group-centered feasibility-level advantage: AiLevel=λL(Li−Lˉ).A_i^{\mathrm{Level}}=\lambda_L(L_i-\bar L).

All other training configurations and the CAD–CAE verifier remain unchanged. We compare Level-only and full HBA at the same training step, evaluated on 3,373 held-out tasks.

Table R1. Level-only vs. HBA.

MethodValid CAD ↑BBox ↑CAE ↑Physics ↑Joint ↑
Level-only98.6491.2596.4452.8651.65
HBA (Ours)99.1491.2596.9257.4055.71

Table R2. Design quality on the 1,472 jointly solved tasks (%).

MethodNormalized Mass ↓Stress Margin ↑Displacement Margin ↑
Level-only83.6425.3830.19
HBA (Ours)79.2326.7736.15

These results demonstrate that PeerRank improves Joint feasibility by 4.06 % points while producing lighter designs with larger physical margins. This confirms the benefit of fine-grained within-level feedback beyond feasibility-level ranking alone.

AuthorsOpenReview ↗

Verification Protocol

3. Geometric-semantic verification and loading consistency.

(1) Geometric-semantic verification. Our verifier distinguishes solid-level geometric checks from program-level semantic checks. Executed CadQuery programs are exported to STEP and inspected through Gmsh, requiring a single volume, non-degenerate geometry, and bounding-box compliance. For semantic requirements, we extract five broad geometric capabilities from the prompt and identify corresponding CadQuery operations in the generated program, as illustrated below.

Geometric capabilityPrompt keywords (examples)CadQuery operations (examples)
Curved profilesCylinder, circular, arc.circle(), .revolve(), .cylinder()
Material removalHole, hollow, drill.hole(), .cut(), .shell()
Polygonal profilesHexagon, triangle, L-shaped.polygon(), .polyline(), .lineTo()
Rectangular profilesRectangle, block, box.rect(), .box(), .polyline()

The listed keywords and operations are illustrative rather than exhaustive.

We clarify that this is a program-level semantic proxy, rather than exhaustive verification of feature counts, spatial locations, or topological relationships in the executed solid. This limitation also motivates our SFT-to-GRPO training strategy: SFT uses a small, separate subset of reference CadQuery programs to establish a reasonable geometric initialization, while subsequent GRPO training relies on executable CAD–CAE feedback without reference CAD supervision. As discussed in our limitations, enforcing more detailed geometric and engineering requirements remains an important direction for future work. We will explicitly clarify this verification scope.

(2) Fixture and loading consistency. This question highlights a central challenge motivating our standardized CAD–CAE pipeline: changes in geometry or topology alter surface identities and potential attachment locations. Identifying physically meaningful fixture and loading regions from arbitrary descriptions would require additional engineering interpretation or manual annotations, making scalable RL training difficult. We therefore adopt standardized boundary conditions rather than inferring application-specific load locations. As detailed in Appendix A.1, each task specifies a fixed analysis axis. For every generated solid, terminal surface groups are selected by centroid projections along this axis, with the negative group fixed and a constant resultant force applied to the positive group. This maintains consistent loading rules across different geometries without requiring identical surface identities. The convention is also included in the model input, as illustrated in Appendix A.3.

AuthorsOpenReview ↗

Independent Validation

4. Independently rechecked for semantic compliance and refined-mesh responses.

(1) Independent verification. We additionally evaluated all Joint-feasible outputs of Qwen3-4B. For semantic verification, Qwen3.5-27B serves as an independent judge with four inputs: (i) the original geometry description, (ii) the source CadQuery program used for dataset construction (The source program is never provided to the RL policy or used as a reference target during RL training.), (iii) the generated CadQuery program, and (iv) geometric measurements obtained by executing both programs (solid counts, through-holes, voids, face and edge types). The judge identifies required elements present in the source design but missing from the generated design, covering shape class, topology, features, and part relationships.

For physical verification, we repeat FEM analysis with half the original element size, preserving material properties, loads, and boundary conditions.

Table R3. Independent verification of all feasible CAD designs.

MetricCalculationResult
Strict Semantic Compliance ↑No required geometric elements missing71.2%
Refined-mesh Physics Pass ↑All physical limits satisfied after refinement97.0%
Median Stress Change ↓Median relative change in p99 stress0.75%
Median Displacement Change ↓Median relative change in max displacement0.48%
Mean Normalized Mass ↓Mean of mass / mass limit over all feasible designs0.807

The full-set semantic compliance rate is determined by the independent judge. Relative FEM changes are measured against the original-mesh responses. Notably, 97.0% of the designs remain physically feasible after mesh refinement, with median stress and displacement changes below 1%, supporting the numerical robustness of our CAD–CAE verification.

(2) CAE/BBox denominators. The apparent discrepancy arises from the distinction between hierarchical training gates and independently reported evaluation metrics. During training, geometric compliance is required before assigning higher feasibility levels. In Table 1, however, BBox and CAE are evaluated as separate success indicators over the same full set of 3,373 tasks. A valid CAD solid may violate its prescribed bounding-box limits while remaining suitable for finite-element analysis. Therefore, CAE success can exceed BBox compliance.

Reviewer_PcQYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents PhyCAD, a physics-aware text-to-CAD generation framework that incorporates finite element analysis (FEA) feedback into reinforcement learning for large language models. The goal is to generate parametric CAD models that not only meet geometric requirements but also satisfy physical constraints. To achieve this, the authors construct a physics-aware text-to-CAD dataset and develop an automated CAD--CAE verification pipeline. They also propose a hierarchical advantage estimation method for group relative policy optimization (GRPO), which divides generated designs into four levels based on their verification outcomes. Within each level, a peer-ranking mechanism provides more fine-grained feedback for reinforcement learning. The authors evaluate their approach on five open-source language models and compare the results with direct generation from three proprietary models. Experiments on 3,373 test tasks show that PhyCAD achieves better joint satisfaction of geometric and physical constraints.

Critical strengths
  1. The task is practically meaningful. Existing Text-to-CAD methods mainly focus on geometric accuracy while often overlooking physical constraints and real-world engineering requirements.

  2. The proposed Hierarchical Balanced Advantage provides more fine-grained reward signals than standard GRPO, which can be useful for tasks involving multiple constraints.

Critical weaknesses

I think PhyCAD is moving in the right direction, but its current implementation is still quite far from what I would consider a truly useful physics-aware CAD generation system. In my view, physics-aware CAD generation should not simply be about having AI generate a CAD model and then checking whether it satisfies physical constraints. Instead, the system should start from real engineering requirements, use physical principles to actively determine what a part should look like, and iteratively improve the design through simulation. In other words, the model should be able to understand the engineering intent, including where the part will be used, what loads it needs to withstand, and which structural features must remain unchanged. Physical requirements such as stress, stiffness, and weight should directly influence the generated geometry, rather than just serving as evaluation criteria after generation.

From this perspective, the current implementation of PhyCAD is still somewhat trivial. What the paper mainly demonstrates is that properly organizing FEM feedback can improve the constraint satisfaction rate of CAD models generated after GRPO training. However, there is still a clear gap between this result and genuine physics-aware CAD generation capabilities. The experiments show that the generated designs are more likely to pass FEM verification, but they do not sufficiently distinguish whether the model has actually learned physics-related design principles or has simply learned statistical shortcuts from the dataset, or RL process.

Regarding the technical contribution, the combination of SFT and GRPO has already been widely adopted in previous work. The main technical novelty of this paper is the proposed Hierarchical Balanced Advantage. However, similar ideas have already been explored in existing literature (e.g., arXiv:2601.23058v1 and multi-objective reinforcement learning papers such as arXiv:2509.22047v1), which the authors neither discuss nor cite.

Finally, the paper claims to develop an automated CAD--CAE verification pipeline, but it does not provide sufficient details about how the underlying simulation system is actually implemented. If the proposed verifier does not introduce any new FEM algorithms and is essentially a system integration of existing CAD modeling and FEM simulation tools, I think the authors should tone down this claim and make the actual technical contribution of this pipeline clearer.

Questions for authors

please see weaknesses

8 comments
AuthorsOpenReview ↗

Weakness 1. physics-aware generation vs. statistical shortcuts (1/2): Contribution Clarification and Benefits Beyond Supervised Imitation

We thank the reviewer for the critical and constructive feedback. Below, we clarify the scope of our contribution and present additional experimental evidence addressing these concerns.

Weakness 1. physics-aware generation vs. statistical shortcuts

1. Clarification of our contribution

We share the reviewer’s broader goal of systems that understand engineering intent and make reliable design decisions across physical conditions. PhyCAD addresses a concrete, testable question within this broader agenda: how can a CAD generator be trained to jointly satisfy geometric and physical requirements while retaining the freedom to determine a feasible design?

As specified in Sec. 3.1 and Fig. 2, each task provides required geometric features, admissible dimensions, material properties, loading and boundary conditions, and limits on stress, displacement, and mass. Within these requirements, the model determines geometric parameters and generates an executable parametric design. The reference CAD serves as one feasible realization, rather than a uniquely prescribed target. This defines a constrained design problem whose success depends on both geometric compliance and simulated physical performance.

Our focus is on learning this constrained generation capability. Richer engineering-intent understanding and simulation-guided iterative refinement remain valuable directions within the broader research agenda. PhyCAD contributes a concrete, empirically evaluated step toward physics-aware CAD generation, supported by improvements on an explicitly formulated and verifiable design task.

2. Benefits beyond additional supervised imitation

To examine whether the improvement can be reproduced by extending reference-program supervision, we introduce an Additional-SFT baseline. Starting from the same SFT initialization, this baseline is further trained on 59,353 reference CAD programs corresponding to the RL training tasks. During RL, PhyCAD instead receives CAD–CAE verification feedback without reference-program supervision. Both models are evaluated on the same 3,373 held-out tasks under identical evaluation settings.

Table 1. Additional supervised imitation versus verification-driven policy optimization (%).

MethodValid CAD ↑BBox ↑Semantics ↑Physics ↑Joint ↑
Additional-SFT93.0679.6695.3538.5735.58
PhyCAD98.4990.7295.4642.0139.16

PhyCAD achieves higher physical constraint satisfaction than Additional-SFT (42.01% vs. 38.57%) and higher joint feasibility (39.16% vs. 35.58%), corresponding to improvements of 3.44 and 3.58 percentage points, respectively. Although Additional-SFT receives the corresponding reference programs as additional supervision, it does not match the performance achieved through CAD–CAE feedback.

This comparison supports the value of optimizing the generator against the geometric and physical consequences of its designs, beyond extending reference-program imitation. The improvement is observed in physical constraint satisfaction as well as overall joint feasibility.

AuthorsOpenReview ↗

Weakness 1. physics-aware generation vs. statistical shortcuts (2/2): Robustness Under Joint Rescaling of Loads and Limits

3. Robustness/Generalization under joint rescaling of loads and limits

We additionally evaluate the same checkpoints, without further training, at 0.5×, 1.0×, and 2.0× the original load. Stress and displacement limits are recalculated from the corresponding reference-CAD responses using the original 1.1× margin and rounding rule. All evaluations use the same 3,373 tasks and scoring criteria.

Under the linear-elastic assumptions of our benchmark, jointly rescaling loads and response limits approximately preserves normalized feasibility conditions, subject to threshold rounding. This experiment therefore evaluates consistency when approximately equivalent physical requirements are expressed at different numerical scales.

Table 2. Joint feasibility when loads and corresponding response limits are rescaled together (%).

Method0.5×1.0×2.0×
Original SFT21.3521.0521.05
Additional-SFT26.1835.5824.34
PhyCAD39.0839.1638.66

PhyCAD maintains joint feasibility of 39.08%, 39.16%, and 38.66%, retaining its advantage across all three settings. By comparison, Additional-SFT decreases by 9.40 and 11.24 percentage points at 0.5× and 2.0× loads relative to its original-load result. PhyCAD therefore sustains higher joint feasibility across the evaluated numerical scales and exhibits greater consistency than Additional-SFT under this protocol.

These comparisons show that the gains from CAD–CAE-guided policy optimization are not reproduced by the reported extension of supervised imitation and are retained under joint rescaling of loads and limits. They support a concrete contribution toward physics-aware CAD generation: learning to produce executable designs that more reliably satisfy explicit geometric and physical requirements, with improved robustness under the evaluated changes in numerical scale. We regard this verified capability as a meaningful step within the broader goal of engineering-driven CAD design.

AuthorsOpenReview ↗

Weakness 2. Technical Novelty and Comparison with Prior RL Methods

Weakness 2. Technical Novelty and Comparison with Prior RL Methods

We thank the reviewer for pointing out these relevant studies. We will include both references and clarify their relationship to PhyCAD. SFT and GRPO constitute our training framework. Our algorithmic contribution lies in how heterogeneous outcomes from sequential CAD–CAE verification are organized into learning advantages.

(1) Relationship to prior methods

RLRR (arXiv:2601.23058v1) introduces relative ranking, bounded rank-based reward corrections, and hierarchical re-ranking that prioritizes correctness, followed by length and quality. PhyCAD shares the motivation of preserving priority while retaining informative comparisons. Its specific construction separates progress across verification levels from objective-wise comparisons among candidates at the same verification level.

MO-GRPO (arXiv:2509.22047v1) independently standardizes each objective before aggregation, addressing imbalance caused by different reward scales and variances. HBA uses objective-wise ranks within verification levels and combines them with an explicit level advantage, incorporating verification-stage precedence alongside multi-objective comparison.

The specific distinction is therefore the joint use of verification status to determine the comparison groups, relevant objectives, and priority structure.

(2) Specific contribution of HBA

CAD–CAE verification provides different feedback at different stages. Programs that fail CAD construction provide construction-progress signals; valid solids that fail geometric checks provide dimensional and semantic signals; and successfully simulated designs provide physical measurements. Moreover, reducing constraint violations and improving an already feasible design require different objective definitions.

HBA reflects this structure through two components:

  • Between-level progress: Li−LˉL_i-\bar L measures a candidate’s verification level relative to its prompt group.
  • Within-level comparison: A~i\widetilde A_i aggregates objective-wise centered ranks among candidates sharing the same prompt and verification level.

The resulting advantage is

Ai=λL(Li−Lˉ)+λwA~i.A_i=\lambda_L(L_i-\bar L)+\lambda_w\widetilde A_i.

With nonnegative objective weights normalized to sum to one, A~i∈[−1,1]\widetilde A_i\in[-1,1]. Consequently, for candidates from the same prompt group,

\quad\Longrightarrow\quad A_i-A_j\ge\lambda_L-2\lambda_w>0$$ when $\lambda_L>2\lambda_w$. **This preserves verification-level precedence in the training advantages while retaining objective-wise distinctions within each level.** The ablation in Sec. 4.5 supports the practical importance of this organization: hierarchical and non-hierarchical GRPO achieve comparable CAD validity, but the hierarchical formulation achieves substantially higher joint feasibility. This demonstrates the importance of organizing verification feedback beyond simply applying GRPO. ### (3) Additional experimental comparison To complement the methodological comparison, we implemented a CAD-adapted hierarchical RLRR-style baseline and evaluated both methods on the same 3,373 validation prompts. **Table 3. Validation performance of the reported RLRR-style and HBA checkpoints (%).** | Method | Valid CAD ↑ | BBox ↑ | CAE ↑ | Physics ↑ | Joint ↑ | |---|---:|---:|---:|---:|---:| | RLRR-style | 93.45 | 79.99 | 90.07 | 42.45 | 39.93 | | **HBA** | **99.38** | **95.43** | **97.69** | **59.92** | **59.00** | At these checkpoints, HBA achieves 59.92% physical constraint satisfaction and 59.00% joint feasibility, compared with 42.45% and 39.93% for the RLRR-style baseline. These checkpoint results supplement the methodological comparison and the hierarchical versus non-hierarchical ablation reported in Sec. 4.5. We will make the relationship to prior work explicit and sharpen the contribution statement accordingly. **HBA contributes a verification-stage-dependent advantage construction that combines cross-level priority with within-level multi-objective ranking.** This specific organization connects the structure of CAD–CAE verification to policy optimization while building on established ideas in relative-reward and multi-objective learning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.