acceptodds
Under review as a conference paper at ICLR 2027

Semantic-Guided Manifold Search for Physical Inverse Design

Abstract

Physical inverse design (e.g., spatial coil design) aims to identify geometric configurations satisfying target physical responses. We consider settings with expensive black-box simulations and no access to simulator gradients. As a practically important representative task, spatial coil design optimizes 3D conductive paths for induction heating and magnetic-field generation, and faces challenging high-dimensional geometric search under limited simulation budgets. Vanilla Bayesian optimization does not by itself specify how domain-relevant local deformations should be constructed from simulation feedback. We present Semantic-Guided Manifold Search (SemGMS), a closed-loop simulation-driven optimization framework. Rather than relying on large-language models to directly generate geometries, SemGMS distills current simulation feedback and historical trials into structured semantic actions. Its core module, the Geometry Grounding Translator (GGT), converts each semantic action into topology-preserving, bounded local deformation families for graph-represented geometries. A surrogate-guided executor performs Bayesian search strictly within these low-dimensional deformation families. Joint geometric feasibility is checked before each simulation. We validate SemGMS on SimCoil-Bench, a unified benchmark for spatial-coil inverse design encompassing welding-loss, magnetic-field, and hybrid multi-objective scenarios. Across five-seed evaluations, SemGMS achieves the lowest mean spatial uniformity errors among the compared methods on all three tasks while maintaining competitive response magnitude.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Large Language Models, Black-Box Optimization, Physical Inverse Design, Geometric Optimization, Bayesian Optimization, Scientific Machine Learning

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_1vJJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed a three-stage pipeline for coil optimization, in which an LLM outputs a semantic command that is translated into a 4-dimensional search subspace, and traditional BO is then used for numerical optimization. I like the idea of bridging the LLM and BO with a search direction (subspace) that leverages the strengths of both, and of mimicking the traditional human-in-the-loop design process by incorporating expert knowledge. My concern then is how general this methodology is, since the GGT is highly manually configured and embedded with domain knowledge. The methodology is also quite close to other publications (e.g., LGBO, ICLR 2026), so a clearer comparison should be provided.

Critical strengths
  1. The idea of using LLM and BO for semantic decision and numerical execution, respectively.
  2. Clear and fair baseline and ablation setup.
Critical weaknesses
  1. The proposed GGT is too specific to the coil case, whereas the paper claims to present a general methodology. There is a lack of clarity about how much the pipeline relies on the proposed GGT (related to Q1), making it hard to distinguish the contributions of the LLM and the representation.
  2. Lack of recent similar works in baselines (e.g., LGBO, LMABO, etc., both in ICLR 2026)
Questions for authors
  1. How much the pipeline relies on the proposed GGT. Especially as reported in Table 2, GGT design will greatly affect the final outcome (both in loss and unif.). What I read from this is that the gain in semGMS is due to a good subspace setup strategy rather than to introducing LLM semantic guidance. What is your comment on this?
  2. Suppose we admit the GGT to be a routine. Is there any non-LLM way to determine the searching manifold (e.g., an online surrogate to predict h = (o,kappa,eta,beta) with expectation improvement that can be naturally integrated into the BO framework)? How is this compared to an LLM planner?
  3. What is the difference between SEMGMS and the recent baselines? This helps to coordinate the contribution.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1vJJ (1/2)

Thank you for your thoughtful comments and questions. We respond to each of your comments below.

W1, Q1, Q3. The proposed GGT is too specific to the coil case, whereas the paper claims to present a general methodology. There is a lack of clarity about how much the pipeline relies on the proposed GGT (related to Q1), making it hard to distinguish the contributions of the LLM and the representation.

How much the pipeline relies on the proposed GGT. Especially as reported in Table 2, GGT design will greatly affect the final outcome (both in loss and unif.). What I read from this is that the gain in semGMS is due to a good subspace setup strategy rather than to introducing LLM semantic guidance. What is your comment on this?

Suppose we admit the GGT to be a routine. Is there any non-LLM way to determine the searching manifold (e.g., an online surrogate to predict h = (o,kappa,eta,beta) with expectation improvement that can be naturally integrated into the BO framework)? How is this compared to an LLM planner?

A1.

  • GGT and the planner contribute at different stages. GGT implements coil-specific deformation modes and feasibility checks. Table 2 shows that these design choices matter. Fixing both focus and range raises the uniformity error from 3.62% to 5.39%, even as welding loss increases from 167.07 to 170.00. The conclusion (§6) specifies 3D spatial-coil inverse design as the application evaluated here.
  • We changed the planner while holding the representation fixed. The LLM and a deterministic rule-based planner share the evidence summaries, action library, GGT, numerical executor, initial FEM observations, and evaluation budget. With symmetric bounds, the LLM planner achieves 158.81 welding loss and 4.34% uniformity error, compared with 135.16 and 6.57% for the rule-based planner. The LLM planner also improves both metrics under sign-restricted bounds. This supports a contribution from LLM action selection with the same geometric representation (§5.3 and Appendix E.5).
  • Expected improvement over actions is also possible. A surrogate for action selection could predict the improvement from executing an action on the current coil under a fixed coefficient-search policy. The empirical comparison in Table R1 evaluates the non-LLM rule-based planner.

Table R1. Matched planner and direction comparison on WLO.

ConfigurationEvidence summariesCoefficient boundsLoss ↑Loss Unif. (%) ↓
LLM plannerDeterministicSymmetric158.814.34
LLM plannerDeterministicSign-restricted153.954.15
LLM-free rule plannerDeterministicSymmetric135.166.57
LLM-free rule plannerDeterministicSign-restricted134.217.44
SemGMS (reference)LLMSymmetric167.073.62

W2. Lack of recent similar works in baselines (e.g., LGBO, LMABO, etc., both in ICLR 2026)

Q3. What is the difference between SEMGMS and the recent baselines? This helps to coordinate the contribution.

A2.

  • Thank you for suggesting these baselines. We added LGBO and LMABO to the WLO comparison in Appendix E.7. SemGMS obtains the highest welding loss and lowest uniformity error among these three results.
  • Role of the LLM. LGBO uses LLM preferences to shift the GP surrogate mean. LMABO uses the LLM to select the acquisition function. In SemGMS, the LLM proposes an action that GGT converts into a local search space of coordinated coil deformations. The numerical executor then selects the deformation coefficients. We added this distinction in the Introduction (§1).

Table R2. WLO comparison with recent LLM-guided BO baselines.

MethodLoss ↑Loss Unif. (%) ↓
LGBO140.656.04
LMABO126.724.53
SemGMS (Ours)167.073.62

We hope our responses address your questions and concerns. We would be happy to discuss any of these points further with you.

I thank the authors for their additional experiments that addressed some of my concerns. I appreciate the matched planner and direction comparison results, which explicitly split the contribution of the LLM planner and LLM-provided evidence, and the baseline test that includes more methods.

However, I still find the results too specific to the coil case. The authors didn’t make much methodological contribution to the machine learning community. They demonstrated that LLMs can assist in optimization with semantic guidance, and that this works in a specific engineering area. Nevertheless, I don’t think a GGT that defines a manifold is common in most complex engineering problems, so applying the proposed methods to other use cases would be difficult. I therefore believe it would fit better in a journal than in ICLR and consequently keep my score (=2).

In addition, a minor concern is that all experiments here are with only one seed. My experience in similar tasks is that optimization is quite sensitive to the seed due to multimodality, so a multi-seed test would be recommended.

AuthorsOpenReview ↗

Response to Reviewer 1vJJ (2/2)

Thank you for your feedback on the additional experiments. We appreciate your questions about the methodological contribution, its applicability beyond coils, and sensitivity to random seeds. 

  • 1. Why this problem matters. Spatial-coil design is a common electromagnetic inverse-design problem in applications such as induction heating and magnetic-field generation. It exemplifies a broader challenge in AI for Science: optimizing high-dimensional designs with a limited budget of expensive black-box simulations and no access to simulator gradients. A small change to one part of a coil can affect the response at several targets, and each candidate is expensive to evaluate. Full-coordinate search must explore many coupled variables under a tight simulation budget, and a fixed parametric design can omit the local changes needed to improve a weak response. SemGMS (our method) uses simulation feedback to identify where and how to edit the coil, translates that edit into a small set of coordinated deformations, and uses Bayesian optimization to determine how much to move. This lets the search adapt its parameterization to the physical response observed in each round. 
  • 2. Methodological contribution and scope. The contribution is to use semantic guidance to construct the numerical search space itself. LGBO uses LLM guidance to adjust the surrogate mean, and LMABO uses it to select the acquisition function. SemGMS instead grounds a semantic action into a low-dimensional family of candidate designs. Proposition 1 (§4.4) gives a conditional local descent result for this construction. If the selected family contains a useful descent direction and enough feasible room to move, the objective improves when the available decrease exceeds the executor’s finite-budget search error. The same approach could apply wherever meaningful edits can be described semantically and translated into feasible changes to design parameters. A new domain would require its own translation rules. Our task covers three spatial-coil design scenarios with distinct objectives and target geometries. 
  • 3. Multi-seed evidence. The main evaluation compares SemGMS with seven baselines over five random seeds on each of the three tasks (§§5.1–5.2, Table 1). Each seed uses a different initial coil, shared by all methods within that seed, and each run has a budget of 200 simulator evaluations. SemGMS achieved the lowest mean uniformity errors across all three tasks. For example, its WLO uniformity error averaged 3.96% with a standard deviation of 0.29 percentage points, and it achieved the lowest welding-loss uniformity error on HFLO in all five runs. Appendix E.1 provides the multi-seed analysis, and Appendix E.3 (Table 11) reports the five-seed WLO comparison with Full-Space GP and Parametric BO.  Thank you again for your careful reading and thoughtful questions. We appreciate the opportunity to discuss these points further.
Reviewer_7KPLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a multi-stage framework for inverse design of coils for inductive heating (welding) and magnetic-field generation. The approach uses an LLM to propose coil design changes which are projected into a feasible geometric modification and evaluated using a surrogate objective function. The approach is evaluated in the aforementioned application areas in simulation.

Critical strengths
  • The approach seems to be reasonably composing LLMs and Bayesian optimization, though important details are missing to fully judge the approach.
  • The paper provides a good set of experiments to evaluate the performance of the proposed approach including ablations.
Critical weaknesses
  • For the general machine learning audience, the applications such as welding are not well enough explained to make the experiments comprehensible.
  • Several terminology is not introduced properly. E.g., "welding-loss", "spatial uniformity error", "response magnitude", "action-conditioned geometric repearametrization", "DSL", "clearance or curvature violation", "tube", "typed operator", "typed DSL translation map", "manifold", "manifold patch", "embedding space", "operation, focus, direction, and coarse range". Especially the latter seem to be central to the method, but it is unclear what the domains of these variables are and what the variables mean semantically.
  • What is an "embedding P_t" ? what is embedded and how?
  • How is V_fix chosen? How many graph nodes are actually optimized in the experiments? How many are fixed and why are those fixed?
  • \phi_mean lacks a proper description of how it is implemented.
  • How is "N_t" defined in Eq 7 ?
  • The theoretical analysis in Sec 4.4 stands without context. It is not clear what is shown/analyzed and how this contributes to the presentation of the paper.
  • Which surrogate function approximation method was used ? GPs ?
  • Why not perform Bayesian optimization directly on the "action space" operator, focus, direction and coarse range?
  • The key working principles of the baseline methods are not well explained and it is unclear how the proposed method differs and why it performs differently. How were these methods adapted for the evaluated tasks ?
  • Sec 5.2 states short conclusions that the proposed stages of the method would be demonstrated by the results, but this is not true. They would need to be tested in isolation, e.g., using an ablation.
  • l. 455, what does it mean that "SGE solves their magnitude" ?
  • l. 461 mentions and ablates "domain priors", but these are not defined prior in the method description. Please clarify what they are.
Questions for authors

The main issues of this paper are in the presentation, e.g., definition of the terminology and proper introduction of the applications, and also in the description of the experimental evaluation and the baselines. Please take special focus in the author response on the issues mentioned in "Weaknesses" related to these points.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7KPL (1/3)

We appreciate the care you took in reading our manuscript and the suggestions you offered. Below, we respond to your comments point by point.

W1. For the general machine learning audience, the applications such as welding are not well enough explained to make the experiments comprehensible.

W2. Several terminology is not introduced properly. E.g., "welding-loss", "spatial uniformity error", "response magnitude", "action-conditioned geometric repearametrization", "DSL", "clearance or curvature violation", "tube", "typed operator", "typed DSL translation map", "manifold", "manifold patch", "embedding space", "operation, focus, direction, and coarse range". Especially the latter seem to be central to the method, but it is unclear what the domains of these variables are and what the variables mean semantically.

A1. Thank you for pointing out where more context would help readers.

  • Welding and physical metrics (§5.1 and Appendices C.2–C.3). We explain how an alternating current in the coil induces heating in the metal tubes to be joined. Welding loss is the power dissipated as heat at a target surface. In WLO, response magnitude measures heating power at the weakest upper and lower weld surfaces, and uniformity error measures its relative variation across tube pairs. The two goals are to deliver more power to the least-heated regions and to distribute that power more evenly.
  • Reparameterization and embedding space (§1 and §3). The embedding space holds the physical 3D coordinates of all coil control points. An action defines a smaller search space by mapping a few deformation coefficients to full coil geometries around the current design. We now explain this connection at the first use of “action-conditioned geometric reparameterization.”
  • DSL and action fields (Figure 2, §§4.1–4.2, and Appendix B.2). A domain-specific language (DSL) is a language designed for a particular problem domain, with notation and abstractions tailored to the tasks in that domain (van Deursen et al.). Figure 2 spells out domain-specific language (DSL), and we now expand it in §4.1. A typed operator specifies its argument types and allowed values. The DSL translation map validates these fields and turns accepted clauses into action tuples. Table 6 and Appendix B.2 already specify the domains and meanings of the four fields. The example in B.2 selects Translate along x (operation), the upper coil layer (focus), the positive direction (direction), and a medium displacement range (coarse range). We added a pointer to the table and example in §4.1.
  • Feasibility (Appendix C.4). The existing distance checks define clearance as the gap between the conductor and nearby objects or nonadjacent coil segments. We now explain that a curvature violation refers to an excessively sharp bend, where the interior angle between neighboring segments is too small. Manifold terminology (§4.2 and Appendix A.3). For a given action, we vary the coefficients of its deformation modes to generate candidate coil shapes. After clipping or rejecting infeasible candidates, the set of feasible shapes we obtain is the local deformation family. In a small neighborhood where the coefficients can vary freely and every resulting shape passes the feasibility checks unchanged, the coil coordinates vary smoothly along the deformation modes. We call this smooth local region a manifold patch (Lee, Introduction to Smooth Manifolds). We clarify this distinction after Eq. (5) in §4.2. Appendix A.3 shows how the patch lies in the affine space through the current coil spanned by the deformation modes.

W3. What is an "embedding P_t" ? what is embedded and how?

A3.

Section 3 defines P as the matrix of coil control-point positions. At iteration t, P_t is the n × 3 matrix of current 3D control-point coordinates for the fixed coil graph G. Each row gives the physical (x,y,z) position of one graph vertex, and the edges specify which points are connected. To avoid ambiguity with encoder outputs, we have revised the wording in §3 to describe P_t as a 3D geometric embedding.

AuthorsOpenReview ↗

Response to Reviewer 7KPL (2/3)

W4. How is V_fix chosen? How many graph nodes are actually optimized in the experiments? How many are fixed and why are those fixed?

A4. The fixed nodes are the power-interface control points on the coil's inlet and outlet paths. Their positions are set by the electrical connections. All other nodes belong to the optimizable coil body. This gives 11 movable control points in the default experiments and 4, 11, and 17 in the one-, two-, and three-layer WLO settings. Appendix C.1 describes the interface paths and the default coil body, and Appendix E.2 gives the layer-specific counts. We now state the selection rule for the fixed set in §3.

W5. \phi_mean lacks a proper description of how it is implemented.

A5. Do you mean the evidence-distillation map in the equation E_t = ϕ_mem(H_t) in §4.1? Appendix B.1, Stage 2 describes its implementation. An LLM updates the previous case memory using the ten most recent verified evidence cards and returns a structured summary of recurring outcomes, failed patterns, and search preferences. Appendix G.3 gives the prompt and a concrete output.

W6. How is "N_t" defined in Eq 7 ?

A6. In §4.1, immediately after defining the evidence memory, the paper states that “N_t is the number of simulator-evaluated designs in memory.”

W7. The theoretical analysis in Sec 4.4 stands without context. It is not clear what is shown/analyzed and how this contributes to the presentation of the paper.

A7. The analysis asks a question that motivates the method. Can restricting the search to a low-dimensional deformation family still yield a better coil? Under Proposition 1's assumptions, the objective improves in that iteration whenever the decrease available within the chosen family exceeds SGE's finite-budget search error. This connects the quality of the search space selected by SAP and GGT to the quality of SGE's numerical search. We added this context at the start of §4.4.

W8. Which surrogate function approximation method was used ? GPs ?

A8. Yes. The main experiments use a Gaussian process, implemented as BoTorch SingleTaskGP, with qLogExpected Improvement (§5.1). Appendix E.5, Table 14 compares it with Bayesian linear regression (BLR) and deep kernel learning (DKL) while holding the planner, grounding, and acquisition procedure fixed.

Table R1. Surrogate comparison on WLO (paper Table 14).

SurrogateLoss ↑Loss Unif. (%) ↓
Bayesian linear regression (BLR)129.665.90
Deep kernel learning (DKL)161.104.56
Gaussian process (default)167.073.62

W9. Why not perform Bayesian optimization directly on the "action space" operator, focus, direction and coarse range?

A9.

  • An action specifies a family of edits. For example, a medium positive-x shift of the upper coil layer admits many different displacements. A BO surrogate over action tuples alone would need a coefficient-selection rule to obtain a coil and a simulator response for each tuple. The same tuple can also produce different edits as the current coil changes. SemGMS uses the latest evidence to choose the action and runs BO over the continuous coefficients of the resulting deformation modes (§§4.1–4.3).
  • We tested action selection under matched conditions. The LLM and rule-based planners use the same deterministic evidence summaries, action library, GGT, and BO executor. The LLM achieves higher welding loss and lower uniformity error under both coefficient bounds. We added this comparison in §5.3 and Appendix E.5.

Table R2. Planner and direction comparison on WLO.

ConfigurationEvidence summariesCoefficient boundsLoss ↑Loss Unif. (%) ↓
LLM plannerDeterministicSymmetric158.814.34
LLM plannerDeterministicSign-restricted153.954.15
LLM-free rule plannerDeterministicSymmetric135.166.57
LLM-free rule plannerDeterministicSign-restricted134.217.44
SemGMS (reference)LLMSymmetric167.073.62
AuthorsOpenReview ↗

Response to Reviewer 7KPL (3/3)

W10. The key working principles of the baseline methods are not well explained and it is unclear how the proposed method differs and why it performs differently. How were these methods adapted for the evaluated tasks ?

A10.

  • §5.1 already groups the baselines into three categories. We have added an explanation of how each works on the coil task. Classical optimizers search the coil coordinates, LLM-as-optimizer methods propose coordinate changes from feedback, and LLM-enhanced optimizers use language models within their candidate search. In SemGMS, the LLM selects a local geometric action and the numerical executor searches the grounded deformation coefficients.
  • Appendices D.4–D.5 and Table 7 specify the adaptations, including initialization, candidate generation, and handling of infeasible proposals. The methods share the coil task, objective, feasibility checks, and simulation budget. The LLM-based baselines receive the same domain information and retain their own update logic and output formats.

W11. Sec 5.2 states short conclusions that the proposed stages of the method would be demonstrated by the results, but this is not true. They would need to be tested in isolation, e.g., using an ablation.

A11. §5.3 already tests the three stages separately in Tables 2–4. Table 2 fixes the GGT focus, range, or both. Table 3 varies SGE's query selection or initialization while keeping SAP and GGT fixed. Table 4 changes SAP's inputs or replaces its planner.

W12. l. 455, what does it mean that "SGE solves their magnitude" ?

W13. l. 461 mentions and ablates "domain priors", but these are not defined prior in the method description. Please clarify what they are.

A12–A13.

  • SGE (§5.3). We revised the wording to “SGE selects the numerical coefficients of the grounded deformation modes.” These coefficients determine how much each coordinated deformation contributes to the proposed coil.
  • Domain priors (§4.1). We moved the definition ahead of the ablation. Domain priors are qualitative physical rules included in the LLM prompts. For example, moving a coil segment closer to a weld region can increase local coupling and heating. Verified simulation results take precedence when they differ from the heuristic.

We hope our responses address your questions and concerns. We would be happy to discuss any of these points further with you.

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

The paper proposes SemGMS, a simulation-driven optimizer for spatial-coil inverse design. SAP uses simulator history to select a semantic coil group, axis, direction, and scale. GGT maps this action to a feasible local deformation family, and SGE uses a Gaussian-process surrogate with qLogExpected Improvement to choose deformation coefficients.

The method is evaluated on the ANSYS Maxwell-based SimCoil-Bench, covering welding-loss optimization (WLO), magnetic-field optimization (MFO), and hybrid field-loss optimization (HFLO). Each run uses 200 simulator evaluations and five seeds. SemGMS reports the best mean uniformity on all three tasks, together with module ablations, layer-count experiments, a full-space GP comparison, and LLM-backbone sensitivity.

Critical strengths
  1. The modular design is clear. LLM planning, deterministic feasibility-aware grounding, and numerical coefficient search have separate roles (Sections 4.1–4.3; Appendices B and C.4). This is a sensible way to use an LLM without asking it to directly generate invalid coil coordinates.

  2. The evaluation is substantial for the chosen simulator setting. The paper reports five-seed comparisons on three objectives, GGT/SGE/SAP ablations, layer-count results, a full-space GP control, backbone sensitivity, and optimization trajectories. The uniformity gains are particularly strong on WLO and HFLO.

  3. The paper is relatively transparent about its limitations. It acknowledges the fixed translation schema, LLM latency, dependence on simulator calls, and the lack of physical-system validation instead of claiming global convergence or complete engineering readiness.

Critical weaknesses
  1. The tested implementation is narrower than the claimed framework, and the semantic contribution is not isolated. All reported experiments use only Translate (Appendix B.2). The implementation uses one group-translation mode and up to three hotspot modes (Appendix B.3). The direction token is not a sign constraint because the coefficient box is symmetric; SGE can search both signs (Appendix G.1). Therefore, the experiments mainly test the selection of a hand-designed local subspace. The paper does not include a direction ablation or a deterministic planner using the same evidence and action family. This does not remove the practical value of the GGT+SGE design. It does mean that the current results cannot determine whether the gain comes from LLM semantic planning, the hand-designed semantic groups, or the low-dimensional grounded search space. A same-space LLM-free planner and a symmetric-versus-direction-restricted comparison would substantially strengthen the claim. Otherwise, claims about richer operators and general manifold search should be narrowed to the tested translation-family implementation. This comparison is also important given related LLM+BO and search-space-partitioning methods such as LLAMBO and HOLLM.

  2. The main evaluation protocol does not fully support the attribution and significance claims. SemGMS spends 100 of 200 evaluations on random initialization and 100 on semantic search, while the baselines use different initialization schemes (Table 7). Thus, sharing the initial coil does not imply a matched initial dataset or warm-start budget. The full-space GP control uses matched initialization but only one seed (Table 11). The five methods share each seed's initial coil, but Appendix E.1 applies independent-sample Kruskal–Wallis and Mann–Whitney tests, ignoring the paired/block structure. With only five seeds, the p-values and effect sizes should be interpreted cautiously; HFLO has no significance analysis. In addition, Equation (1) uses an unnormalized max of threshold-referenced losses and a soft-constraint penalty, but the complete thresholds and penalty weights for WLO, MFO, and HFLO are not reported. LLM-call counts, rejected/time-out simulations, latency, and wall-clock cost are also absent. The reported means are encouraging, but the protocol does not yet establish how much of the gain comes from initialization, semantic planning, or numerical execution. A matched-initialization comparison, per-seed results, block-aware statistics, and complete objective/cost configurations are needed.

  3. The evidence and theory support a narrower conclusion than the paper's broad framing. The default study uses ANSYS Maxwell 19.2, one four-pair scene, a two-layer 11-point coil, and five perturbed initial coils. Layer-count and WLO-to-HFLO tests add useful variation, but do not establish transfer to new scenes, topologies, simulators, or physical coils. The conclusion should therefore be limited to the tested Maxwell benchmark unless broader evidence is added. Theoretical Proposition 1 is a useful motivation, but it is conditional on a smooth active branch, locally inactive projection, and a relative-ball condition. The experiments do not measure its alignment score or execution-error term, so the proposition does not explain why the tested semantic families outperform the baselines. This is acceptable as a local analysis, but it should not be presented as empirical support for the general framework.

  4. Minor Presentation Issues Use “local deformation family” consistently when feasibility clipping or rejection makes the feasible set piecewise smooth; “manifold” is too strong for the full feasible set. Correct the typo “lsoss” in Section 5.2.

Questions for authors
  1. What causal role does the direction token have when coefficients can take either sign? Please compare direction-restricted and symmetric coefficient boxes, and add an LLM-free planner with the same evidence and action family.

  2. How many initialization evaluations does each baseline receive? Please provide a matched-100-initialization comparison, per-seed results, block-aware tests, and the complete WLO/MFO/HFLO thresholds and penalty weights.

  3. How many LLM calls, rejected/time-out simulations, and seconds of wall-clock time does one 200-evaluation run require? Do the authors have results on an unseen scene, topology, simulator, or physical coil? If not, please narrow the conclusion to the tested Maxwell benchmark.

3 comments
AuthorsOpenReview ↗

Response to Reviewer HDJT (1/3)

Thank you for taking the time to review our manuscript and share your feedback. We address your questions and concerns below.

W1, Q1. The tested implementation is narrower than the claimed framework, and the semantic contribution is not isolated. All reported experiments use only Translate (Appendix B.2). The implementation uses one group-translation mode and up to three hotspot modes (Appendix B.3). The direction token is not a sign constraint because the coefficient box is symmetric; SGE can search both signs (Appendix G.1). Therefore, the experiments mainly test the selection of a hand-designed local subspace. The paper does not include a direction ablation or a deterministic planner using the same evidence and action family. This does not remove the practical value of the GGT+SGE design. It does mean that the current results cannot determine whether the gain comes from LLM semantic planning, the hand-designed semantic groups, or the low-dimensional grounded search space. A same-space LLM-free planner and a symmetric-versus-direction-restricted comparison would substantially strengthen the claim. Otherwise, claims about richer operators and general manifold search should be narrowed to the tested translation-family implementation. This comparison is also important given related LLM+BO and search-space-partitioning methods such as LLAMBO and HOLLM.

What causal role does the direction token have when coefficients can take either sign? Please compare direction-restricted and symmetric coefficient boxes, and add an LLM-free planner with the same evidence and action family.

A1. In the matched WLO ablation, the LLM planner achieved higher welding loss and lower uniformity error than the rule-based planner under both coefficient bounds, with GGT and SGE held fixed.

  • One operator covers both group shifts and local shape edits. We deliberately use a single DSL operator, Translate. Its group mode shifts a coil segment or layer, while up to three hotspot modes let selected points move by different amounts. Together they can produce layer offsets and approximate local bends or smoothing where the modes support those edits. We clarify this coverage in §4.2 and Appendix B.3.
  • The direction sign guides the next round through feedback. If the LLM suggests +x, symmetric bounds let SGE try either +x or −x. The next round sees which move was actually tested and how the simulator metrics changed. The LLM can use that evidence to keep or revise its earlier suggestion (Appendices B.1 and G.1).
  • The WLO ablation supports LLM planning and bidirectional search. We compared LLM and rule-based planners using the same evidence and executor under both coefficient bounds (Table R1). The LLM achieved higher welding loss and lower uniformity error under either bound. Allowing both signs yielded higher welding loss for both planners and lower uniformity error for the rule planner. With the LLM planner, the sign restriction gave slightly lower uniformity error (4.15% vs. 4.34%). See §5.3 and Appendix E.5.
  • Relation to LLAMBO and HOLLM. SemGMS grounds simulator-driven actions into path-aware deformation families for a numerical executor. LLAMBO uses LLMs for warm-starting, surrogate modeling, and candidate sampling, while HOLLM partitions the search space for LLM sampling.

Table R1. WLO ablation of planning and direction constraints.

ConfigurationEvidence summariesCoefficient boundsLoss ↑Loss Unif. (%) ↓
LLM plannerDeterministicSymmetric158.814.34
LLM plannerDeterministicSign-restricted153.954.15
LLM-free rule plannerDeterministicSymmetric135.166.57
LLM-free rule plannerDeterministicSign-restricted134.217.44
SemGMS (reference)LLMSymmetric167.073.62
AuthorsOpenReview ↗

Response to Reviewer HDJT (2/3)

W2, Q2, Q3. The main evaluation protocol does not fully support the attribution and significance claims. SemGMS spends 100 of 200 evaluations on random initialization and 100 on semantic search, while the baselines use different initialization schemes (Table 7). Thus, sharing the initial coil does not imply a matched initial dataset or warm-start budget. The full-space GP control uses matched initialization but only one seed (Table 11). The five methods share each seed's initial coil, but Appendix E.1 applies independent-sample Kruskal–Wallis and Mann–Whitney tests, ignoring the paired/block structure. With only five seeds, the p-values and effect sizes should be interpreted cautiously; HFLO has no significance analysis. In addition, Equation (1) uses an unnormalized max of threshold-referenced losses and a soft-constraint penalty, but the complete thresholds and penalty weights for WLO, MFO, and HFLO are not reported. LLM-call counts, rejected/time-out simulations, latency, and wall-clock cost are also absent. The reported means are encouraging, but the protocol does not yet establish how much of the gain comes from initialization, semantic planning, or numerical execution. A matched-initialization comparison, per-seed results, block-aware statistics, and complete objective/cost configurations are needed.

How many initialization evaluations does each baseline receive? Please provide a matched-100-initialization comparison, per-seed results, block-aware tests, and the complete WLO/MFO/HFLO thresholds and penalty weights.

How many LLM calls, rejected/time-out simulations, and seconds of wall-clock time does one 200-evaluation run require?

A2.

  • Initialization and simulation budget (§5.1, Table 7). Each baseline follows its prescribed initialization. Methods that need an initial sample set generate one, and methods that can start directly begin from the shared initial coil. Every run uses 200 simulator evaluations, including initialization.
  • Matched WLO comparison (Appendix E.3, Table 11). We extended the control to five seeds. Within each seed, SemGMS and Full-Space GP share the same 100 initial observations, GP, and acquisition function. SemGMS attains higher mean welding loss and lower mean uniformity error than both controls in Table R2.
  • Objective settings (Eq. 1 and Appendix C.3). WLO targets welding loss ≥ 115.0 and loss-uniformity error ≤ 8.0%. MFO targets field strength ≥ 95.0 and field-uniformity error ≤ 30.0%. HFLO targets welding loss ≥ 115.0 and both uniformity errors ≤ 8.0%. All three tasks use λ = 0 for the soft penalty. Coils that violate hard constraints are rejected before simulation, equivalent to an objective of +∞.
  • Runtime and LLM usage (§5.4). Maxwell accounts for over 80% of runtime in the WLO backbone study. Across three backbones, SemGMS averaged 3.04 LLM requests per semantic round (about 304 per run) with a 97.69% request success rate (890/911).

Table R2. Five-seed WLO results (mean ± sample SD).

MethodLoss ↑Loss Unif. (%) ↓
Full-Space GP127.98 ± 4.065.70 ± 0.34
Parametric BO149.73 ± 8.868.20 ± 0.09
SemGMS (Ours)166.36 ± 2.523.96 ± 0.29
AuthorsOpenReview ↗

Response to Reviewer HDJT (3/3)

W3Q3. The evidence and theory support a narrower conclusion than the paper's broad framing. The default study uses ANSYS Maxwell 19.2, one four-pair scene, a two-layer 11-point coil, and five perturbed initial coils. Layer-count and WLO-to-HFLO tests add useful variation, but do not establish transfer to new scenes, topologies, simulators, or physical coils. The conclusion should therefore be limited to the tested Maxwell benchmark unless broader evidence is added. Theoretical Proposition 1 is a useful motivation, but it is conditional on a smooth active branch, locally inactive projection, and a relative-ball condition. The experiments do not measure its alignment score or execution-error term, so the proposition does not explain why the tested semantic families outperform the baselines. This is acceptable as a local analysis, but it should not be presented as empirical support for the general framework.

Do the authors have results on an unseen scene, topology, simulator, or physical coil? If not, please narrow the conclusion to the tested Maxwell benchmark.

A3.

  • Experimental scope. The abstract identifies 3D spatial-coil design as a representative physical inverse-design task. In Maxwell 19.2, SimCoil-Bench covers welding-surface heating (WLO), magnetic-field generation in a 3D target volume (MFO), and joint welding-loss and field optimization across evaluation planes (HFLO). The WLO layer study also tests one-, two-, and three-layer coils with 4, 11, and 17 optimizable control points, respectively (§5.1 and Appendices C.2 and E.2). We have also clarified in the conclusion (§6) that the framework targets 3D spatial-coil inverse design.
  • Proposition 1 addresses a local mechanistic question. §4.4 already states, “This provides local motivation.” The proposition assumes a locally smooth objective branch, an inactive feasibility projection, and a local neighborhood within the span of the deformation modes. Its bound predicts descent when the gain along that span exceeds the executor's finite-budget search error. Table 1 reports measured performance: SemGMS achieves the lowest mean uniformity errors across all three tasks and the highest mean welding loss in WLO and HFLO.

W4. Minor Presentation Issues Use “local deformation family” consistently when feasibility clipping or rejection makes the feasible set piecewise smooth; “manifold” is too strong for the full feasible set. Correct the typo “lsoss” in Section 5.2.

A4.

  • We will clarify the terminology after Eq. (5) in §4.2, consistent with Appendix A.3. We use local deformation family for the full grounded set and manifold patch for the smooth local patch where the feasibility map acts as the identity.
  • Thank you for catching the typo. We will correct lsoss to loss in §5.2.

We hope our responses address your questions and concerns. We would be happy to discuss any of these points further with you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Responses

Dear ACs and Reviewers,

Thank you for reading our paper carefully and for your thoughtful questions and suggestions. Your feedback helped us clarify how SemGMS works, strengthen the experimental evaluation.

1. Strengths recognized in the reviews

  • Connecting LLM reasoning and Bayesian optimization through a search subspace. "I like the idea of bridging the LLM and BO with a search direction (subspace) that leverages the strengths of both, and of mimicking the traditional human-in-the-loop design process by incorporating expert knowledge." [Reviewer 1vJJ]
  • Clear design. "The modular design is clear." [Reviewer HDJT]
  • Experimental coverage and ablations. Reviewer 7KPL wrote, "The paper provides a good set of experiments to evaluate the performance of the proposed approach including ablations." Reviewer 1vJJ also noted, "Clear and fair baseline and ablation setup." [Reviewer 7KPL]
  • Strong uniformity across the three tasks. "SemGMS reports the best mean uniformity on all three tasks". [Reviewer HDJT]
  • Transparent discussion of the method's limitations. "The paper is relatively transparent about its limitations." [Reviewer HDJT]

2. Main revisions

In the revised manuscript, blue marks added explanations and clarifications, and purple marks new experiments, analyses, or results.

The Introduction (§1) now explains where LLM guidance enters each method. LGBO uses LLM preferences to shift the GP surrogate mean, LMABO uses the LLM to select the acquisition function, and SemGMS uses a proposed geometric action to construct the local deformation space searched by BO. The Conclusion (§6) specifies 3D spatial-coil inverse design as the application evaluated in this work.

In §3, we clarified that P_t gives the current 3D coordinates of the coil graph, whose connectivity stays fixed. We also explained why the inlet and outlet power-interface points stay fixed. §5.1 and Appendices C.2–C.3 now explain how induction heats the metal tubes, what welding loss and uniformity error measure, and why WLO seeks more heating power at weakly heated surfaces and more even power delivery across tube pairs. Appendix C.3 lists the objective thresholds for WLO, MFO, and HFLO. Appendix C.4 clarifies clearance and sharp-bend terminology.

In §§4.1–4.2, we spelled out DSL as domain-specific language on first use, described the physical rules used as domain priors, and explained how typed action fields are validated and grounded. We pointed readers to the permitted field values and worked example in Appendix B.2. We also distinguished the full local deformation family from a smooth manifold patch where the feasibility map acts as the identity. §4.4 now introduces the question addressed by Proposition 1. Under its assumptions, a low-dimensional action family can yield an improvement when the decrease available within it exceeds the executor's finite-budget search error.

In §5.3 and Appendix E.5, we added a WLO comparison that varies the planner and coefficient bounds while using the same deterministic evidence-summary procedure, GGT, and numerical executor. With symmetric bounds, the LLM planner achieved 158.81 welding loss and 4.34% uniformity error, compared with 135.16 and 6.57% for the deterministic rule-based planner. The LLM planner improved both metrics under sign-restricted bounds as well.

In §5.1, we explained how the baseline optimizers are adapted to the coil task and clarified that their respective initialization evaluations count toward the common 200-simulation budget. In Appendix E.3 (Table 11), we extended the WLO comparison with Full-Space GP and Parametric BO to five seeds. Full-Space GP and SemGMS share the 100 initialization observations within each seed, the GP, and the acquisition function. SemGMS attains 3.96% mean uniformity error, compared with 5.70% for Full-Space GP and 8.20% for Parametric BO.

In Appendix E.7, we added LGBO and LMABO as WLO baselines. Their welding loss and uniformity errors are 140.65 / 6.04% and 126.72 / 4.53%, respectively. The corresponding SemGMS result is 167.07 / 3.62%.

In §5.4, we added cost measurements from the WLO backbone study. Maxwell accounts for over 80% of runtime in the profiled runs. SemGMS averages 3.04 LLM requests per semantic round, with a 97.69% request success rate (890/911) across the three backbones.

Thank you again for your time and for helping us improve the manuscript.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.