acceptodds
Under review as a conference paper at ICLR 2027

VIGOR: Verifiable Industrial Geometry Optimization and Repair in Scene Programs

Abstract

Large language models (LLMs) increasingly generate structured layouts and executable scene programs from natural-language instructions. However, executability does not imply engineering feasibility: an industrial scene program may execute successfully while its final state still violates coupled boundary, collision, clearance, aisle, or structural constraints. We identify this mismatch as the executability–engineering-feasibility gap, where locally plausible placements can remain globally infeasible because corrections to one device may introduce violations elsewhere. To make this gap measurable in a controlled setting, we construct an execution-grounded industrial scene-program dataset with 1,749 constructible scenarios, complete executable programs, and 94 final-state geometric checks per scenario. We further propose VIGOR, which combines supervised scene-program generation with execution-grounded verification and staged geometric refinement. Supervised fine-tuning establishes complete program generation, local projection resolves independently correctable violations, and joint conflict-graph repair coordinates residual multi-object conflicts while preserving structural relations. Experiments show that VIGOR achieves 96.86% Complete Feasibility and an Overall score of 91.20, outperforming the strongest evaluated baseline by 10.42 points in Overall. These results show that, under controlled industrial layout conditions, reliable scene-program generation requires not only executable programs, but also explicit verification and coordinated resolution of coupled engineering constraints.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Industrial Scene Program Generation, Executable Scene Verification, Geometric Constraint Satisfaction, Layout Refinement, Engineering Feasibility, Digital Twin Construction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uEmgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

VIGOR studies industrial scene-program generation, distinguishing program executability from complete engineering feasibility of the executed layout. The authors construct a controlled benchmark with 1,749 constructible conditions, fixed sets of 11 instances, 94 geometric checks per scene, and additional structural checks. The system combines supervised program generation, local geometric projection, and residual conflict-graph-based joint repair while preserving parent and attachment relations. It achieves 96.86% complete feasibility; ablations support complementary roles for SFT and the two repair stages.

Critical strengths
  1. A controlled benchmark with explicit constraints and execution-validated references. Each retained condition has a complete reference program checked for execution, save–reload consistency, geometry, and structure. Interpolation and held-out parameter-combination tests provide a useful experimental basis for studying constraint satisfaction under changing layout conditions.
  2. Substantial empirical gains and evidence of complementary repair stages. Complete feasibility rises from 14.57% with SFT alone to 56.86% with local repair, 81.43% with joint repair, and 96.86% with both. These results support the practical value of staged repair and its complementarity with initial program generation.
Critical weaknesses
  1. The relative advantage of the proposed joint repair is not established. LayoutVLM jointly optimizes geometry and spatial relations; Procedural Scene Programs coordinates objects through shared parameters and structure-preserving program search; Global Graph-Validated Optimization combines relational verification with global–local optimization. These works are cited but not directly evaluated. The ablations establish that joint repair is useful within VIGOR, but there is no comparison that fixes the paper’s SFT model and local-repair outputs while replacing only the joint-repair module with an existing approach. The advantage of this particular repair strategy therefore remains unclear.
  2. Limited structural coverage weakens the benchmark’s case as a major contribution. All conditions share a fixed workspace, asset inventory, attachment structure, and reference operation pattern; variation primarily concerns aisle width, equipment clearance, boundary clearance, and aisle orientation. This controlled design is reasonable, but the tests mainly establish parameter generalization within one scene setting. The paper does not sufficiently establish broader diagnostic value or compare against its reference-construction pipeline or an independent constraint solver, leaving the added value of learned generation in this setting unclear.
  3. The core repair procedure is insufficiently specified for reproduction. Section 3.4 provides an objective and search framework, but does not adequately define successor generation, coordinated multi-object adjustments, or the actual search configuration. Search budgets and runtime are also unreported. These choices directly affect repair success and computational cost, making it difficult to reproduce the main results or assess the practical cost of the feasibility gains.
  4. Functional-quality evaluation lacks verifiable definitions, weakening the interpretation of the overall performance advantage. Appendix G.2 lists six scoring dimensions and their aggregation, without specifying their scoring rules, thresholds, or engineering basis. Calculated from Table 2 and Eq. (15), functional quality accounts for approximately 69% of the Overall gain over the strongest baseline (about 7.15 of 10.42 points). This omission therefore materially affects the main comparison. The reported LLM–human agreement concerns separate OU/SO visual ratings and does not directly validate these functional scores.
Questions for authors
  1. With the paper’s SFT model and local-repair outputs fixed, are results available when only the joint-repair module is replaced by existing multi-object optimization or structure-preserving repair methods? Under the same adapted constraints and comparable computational budgets, what advantage does the proposed repair provide in complete feasibility, structural preservation, or runtime?

  2. What diagnostic value and scope of applicability does the fixed-scene parameter benchmark establish? Are comparisons with the reference-construction pipeline or independent constraint solvers available to demonstrate the added value of learned generation in success rate, cost, or functional quality?

  3. Please specify how N(z) generates successors, how multi-object adjustments are coordinated, and how the conflict graph is used and updated. What beam width and stopping budgets were used, and what are the per-scene runtime or verifier-call counts? How do feasibility and cost change with the search budget?

  4. Please provide the explicit rules, thresholds, and engineering basis for all six functional scores in Appendix G.2, together with evidence of agreement with human functional assessments or practical requirements. Given that functional quality accounts for approximately 69% of the Overall gain, how do these definitions and validation support the interpretation of the overall advantage?

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer uEmg (Part 1/3)

We thank you for your detailed feedback on joint repair, the scope of the fixed-scene benchmark, search reproducibility, and the validity of the functional-quality evaluation.

Q1. Joint repair with a fixed generator and local-repair stage (W1&Q1).

A1: Joint repair raises Complete Feasibility by 10.00 percentage points over LayoutVLM, the strongest alternative in this comparison. All four pipelines start from the same fine-tuned generator and local-repair outputs. When local repair leaves a conflict, we replace only J with LayoutVLM, Procedural Scene Programs, or the graph-validated optimizer (GPFS).

MethodComplete FeasibilityStructural PreservationAvg. Repair Runtime
LayoutVLM86.86%99.71%2.711 s
Procedural Scene Programs70.00%99.71%1.402 s
GPFS67.71%99.71%1.992 s
VIGOR (J)96.86%99.71%4.292 s

All four methods used the same structural constraints and preserved the required structure in 99.71% of cases. The gain over LayoutVLM cost 1.581 s of additional mean repair time (4.292 vs. 2.711 s).

Q2. Diagnostic value of a fixed workspace under varying layout constraints (W2&Q2).

A2: Fixing the 7.5 × 7.0 m workspace and 11 required instances lets us test whether a generated program can satisfy all 94 geometric checks while preserving the required structure. The compositional split holds out combinations of aisle width and equipment clearance while retaining their individual values in training (Appendix C). In Table 3, the fine-tuned generator scores 92.79 on Geometry but achieves only 14.57% Complete Feasibility. With local and joint repair, feasibility reaches 96.86%.

(1) Reference constructor. Of 2,754 candidate conditions, the benchmark retained the 1,749 for which the constructor produced a verified layout. Its 100% test feasibility follows from this selection. Functional Layout measures the quality of the arrangement on those same conditions.

MethodComplete FeasibilityFunctional LayoutAvg. Runtime
Constructor100.00%70.9364.440 s
VIGOR (Ours)96.86%72.3990.839 s

The constructor is faster and solves every retained test condition. VIGOR solves 96.86% and scores 1.46 points higher on mean Functional Layout (72.39 vs. 70.93).

(2) Variation within the workspace. We evaluated 24 conditions varying aisle width, equipment clearance, boundary clearance, and aisle orientation.

MethodComplete FeasibilityFunctional LayoutProcess Chain
Constructor24/24 (100%)72.2041.39
VIGOR24/24 (100%)73.0758.18

Both methods satisfied all constraints on all 24 conditions. VIGOR led by 16.79 points on Process Chain (58.18 vs. 41.39) and 0.87 on Functional Layout. This experiment tests the arrangement of the same equipment under different spatial requirements. Transfer to different factory footprints and asset inventories remains open.

AuthorsOpenReview ↗

Author Response to Reviewer uEmg (Part 2/3)

Q3. Joint-search moves, stopping rules, and computational cost (W3&Q3).

A3: Joint search is invoked only for executable layouts that remain infeasible after local repair (Eq. 4). Across all scenes, the L/J stages averaged 4.132 s per scene, excluding initial generation. On scenes that invoked J, that stage averaged 4.292 s, with a median of 319.5 verifier calls.

(1) Successors. J builds a graph connecting devices involved in residual conflicts and selects at most two connected components per state, prioritizing the most severe conflicts. Single-device translation, pair separation, coordinated displacement, and constrained rotation generate successors. Each successor is re-executed and verified, preserving previously satisfied parent and attachment relations. The graph is rebuilt from the resulting conflicts.

(2) Beam and stopping rules. Each expansion generates at most 60 candidates. The beam retains up to 16 states under Eq. (7), which first ranks layouts by violation count, then total and maximum normalized deficit. Functional score, number of modified devices, displacement, and a deterministic tie-breaker follow. Search stops when it finds a feasible layout, exhausts the candidates, or reaches a limit of 12 levels, 2,000 verifier calls, or 60 s. A replacement must improve the input objective and preserve attachments.

(3) Budget sensitivity. On the same 64 test tasks spanning different task types, raising the cap from 500 to 1,000 calls yielded two additional feasible layouts. Doubling it again yielded one more.

J Search BudgetComplete FeasibilityAvg. Repair RuntimeAvg. Validator Calls
50050/642.680 s198.45
1,00052/644.104 s299.98
2,00053/645.658 s410.70

Across the full budget range, feasibility rose from 50/64 to 53/64 and mean repair time increased by 2.978 s. We will add the move set, stopping rules, and timing breakdown to Section 3.4.

AuthorsOpenReview ↗

Author Response to Reviewer uEmg (Part 3/3)

Q4. Definition, engineering basis, and sensitivity of the six functional scores (W4&Q4).

A4: Functional Layout averages six distance-based scores. We checked the Process Chain score against an obstacle-avoiding route length and tested all six scores for sensitivity to their thresholds.

(1) Scoring rules. All distances are horizontal and measured in metres. For process, tool, rack, container, and sign placement, the evaluator assigns full credit on a target interval and tapers linearly on either side.

\begin{cases} 0, & d\le a\\ \text{or}\\ d\ge e, \\\\ (d-a)/(b-a), & a<d<b, \\\\ 1, & b\le d\le c, \\\\ (e-d)/(e-c), & c<d<e. \end{cases}$$ With $C(d)=\operatorname{clip}((d-1.5)/1.5,0,1)$, the six scores are given by $$S_{\mathrm{process}} = 100T(d_p;1.2,1.5,2.0,3.0)$$ $$S_{\mathrm{tool}} = 100T(d_t;1.2,1.5,1.9,2.4)$$ $$S_{\mathrm{rack}} = 100T(d_r;0,2.0,3.0,4.0)$$ $$S_{\mathrm{container}} = 100T(d_c;0,1.0,2.0,3.0)$$ $$S_{\mathrm{desk}} = 50[C(d_{D1})+C(d_{D2})]$$ $$S_{\mathrm{sign}} = 50[T(d_{S1};0,1.5,3.0,4.5)+T(d_{S2};0,1.5,3.0,4.5)]$$ Here d_p measures the distance from the combined bounding region of the kit-tray station and automated guided vehicle (AGV) to the assembly table. d_t is the distance from the tool cart to the kit-tray station. d_r and d_c measure rack and container distances to the assembly table. The desk and sign each have two distances, one to the AGV and one to the robot proxy. The per-task score is the arithmetic mean of the six scores. Invalid or unscorable tasks receive zero and remain in the method average (Eq. 14). (2) **Engineering basis and route-length check.** CCOHS guidance on materials flow and workspace layout motivates short transfers and adequate operating space. OSHA 1910.176(a) addresses aisle clearance, HSE guidance addresses pedestrian–vehicle separation, and Directive 92/58/EEC addresses sign placement near hazards and in visible locations. These principles motivate the choice of measures. The numerical breakpoints are benchmark parameters, and sign distance measures proximity alone. On tasks where both methods produced feasible layouts, the mean obstacle-avoiding process-chain path was **2.360 m for VIGOR and 3.522 m for DeepSeek-V4.1-Flash + L/J**. Route length is measured separately from the six horizontal-distance scores. (3) **Contribution to Overall and robustness.** Eq. (15) gives one quarter of Overall to Q, the mean of Functional Layout, BQ-FQ, Interp-FQ, Comp-FQ, and Tight-FQ. Against the original Table 2 baseline, DeepSeek-V4.1-Flash without L/J, VIGOR gains **28.600 points on Q**. This contributes **7.150 of its 10.416-point Overall gain (68.64%)**. We varied each score’s thresholds by ±10% separately, then all six together, giving 14 perturbations. | **Metric** | **Original Improvement** | **Improvement under 14 Perturbations** | | ------------ | -----------------------: | -------------------------------------: | | Overall | +10.416 | **+9.083 \~ +11.075** | | Functional Q | +28.600 | **+23.268 \~ +31.235** | VIGOR retained an Overall lead of at least **9.083 points** across all 14 settings. We will add the score definitions, route-length check, and sensitivity results to Appendix G. We would be grateful if you would **reconsider your evaluation** in light of the controlled optimizer comparison and the robustness of the functional-quality results.
Reviewer_FpEROpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the generation of feasible industrial scenes for virtual factories. VIGOR combines an LLM fine-tuned to generate JSON scene programs with local and joint geometric repair. Local repair corrects individual boundary, aisle, and clearance violations, while joint repair coordinates changes to objects involved in shared conflicts.

VIGOR outperforms the evaluated baselines in complete feasibility and overall score. Ablations show that combining fine-tuning with both repair stages yields better feasibility than partial configurations.

Critical strengths
  • The paper addresses a relevant problem: executable scene programs may still violate important spatial and structural constraints.

  • It introduces an execution-grounded dataset of 1,749 constructible scenarios with explicit feasibility checks.

  • The reported results are strong within this controlled setting, and the ablations demonstrate the complementary benefits of local and joint repair.

Critical weaknesses
  • The paper does not compare VIGOR with its deterministic reference-construction pipeline or a standard constraint solver. This makes it difficult to assess the practical advantage of learned generation.

  • DeepSeek-V4.1-Flash achieves substantially higher complete feasibility than the SFT-only model. Although the paper tests repair on the untuned base model, it does not establish whether task-specific fine-tuning offers an advantage over a stronger model followed by the same repair pipeline.

Questions for authors
  1. What advantage does VIGOR provide over the deterministic reference-construction pipeline or a standard constraint solver on the same test conditions, considering feasibility, functional quality, and runtime? Does reference construction use any information unavailable at deployment?

  2. How does the strongest structurally competent baseline perform with the same local–joint repair pipeline? This comparison would help separate the contribution of fine-tuning from that of geometric repair.

  3. Could the authors specify the local and joint candidate moves, beam width, stopping budgets, deficit normalization, six functional scoring functions, and baseline format adapters? These details are important for reproducing the method and interpreting the comparisons.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer FpER (Part 1/2)

We thank you for your constructive questions about the practical advantage of learned generation, the contribution of task-specific fine-tuning, and the details needed for reproducibility.

Q1. Practical advantage over deterministic construction and retrieval (W1&Q1).

A1: The reference constructor can run at deployment. It uses the workspace and asset geometry, the task constraints, and predefined placement rules. Its 100% feasibility follows from how the benchmark was built. Only 1,749 of 2,754 candidate conditions passed construction and validation (Appendix C), and all test tasks come from that retained set. We compared it with VIGOR and two nearest-neighbor baselines on the same test tasks using the same final validator.

MethodComplete FeasibilityFunctional LayoutAvg. Runtime
Constructor100.00%70.9364.440 s
Retrieval-only16.29%70.260.054 s
Retrieval+L/J88.29%71.481.079 s
VIGOR (Ours)96.86%72.3990.839 s

The constructor is faster than VIGOR and solves every retained test task. VIGOR achieves the highest mean Functional Layout score, 72.39 versus 70.93 for the constructor and 71.48 for retrieval followed by repair. Against repaired retrieval, the gains are 15.98 points on Process Chain and 7.91 on Tool/Kitting. Retrieving a layout from the prebuilt library takes 0.054 s on average, rising to 1.079 s with repair.

Q2. Contribution of task-specific fine-tuning under the same repair procedure (W2&Q2).

A2: Fine-tuning raises Overall from 24.03 to 91.20 on the same backbone. With local and joint repair (L/J) fixed, untuned Qwen3-VL-8B-Instruct yields no feasible layouts. Fine-tuning it on the reference programs raises Complete Feasibility to 96.86% (Table 3). We also replaced the initial generator with three other models, keeping the test tasks, repair procedure, search budget, and final validator fixed.

MethodComplete FeasibilityFunctional QOverall
DeepSeek-V4.1-Flash + L/J99.14%40.1784.82
GPT-5.5 + L/J62.86%36.6971.85
Ministral-3-8B + L/J0.00%29.1229.76
VIGOR96.86%68.6291.20

DeepSeek-V4.1-Flash with repair has the highest feasibility (99.14%). VIGOR leads it by 28.45 points in Functional Q and 6.38 in Overall under the same repair procedure. Functional Q combines mean Functional Layout with the bottom-quartile, interpolation, compositional, and tight-condition scores (Appendix G.3). This accounts for VIGOR’s 68.62 here and its 72.39 mean Functional Layout in A1.

AuthorsOpenReview ↗

Author Response to Reviewer FpER (Part 2/2)

Q3. Repair settings, scoring functions, and baseline format adapters.

A3: (1) Search moves and budgets. Local repair LL projects devices inside the boundary, moves them out of the reserved aisle, and corrects equipment clearances. Its limits are 40 iterations, 800 validator calls, and 30 s. Joint repair JJ builds a conflict graph from the current layout and proposes single-device translations, two-device separations, coordinated displacements, and constrained rotations. Beam width is 16, maximum depth is 12, and each expansion generates at most 60 candidates. The limits for JJ are 2,000 validator calls and 60 s. Search stops at feasibility, candidate exhaustion, or the first applicable limit. Eqs. (6)–(8) specify candidate acceptance and ranking.

(2) Violation deficits. With lengths in metres, the normalized boundary, aisle, and equipment-clearance deficits are

Dboundary=gbmax⁡(b,0.001)D_{\mathrm{boundary}}=\frac{g_b}{\max(b,0.001)}

Daisle=owmax⁡(w,0.001)D_{\mathrm{aisle}}=\frac{o_w}{\max(w,0.001)}

Dclearance=max⁡(0,c−dxz)max⁡(c,0.001)D_{\mathrm{clearance}}=\frac{\max(0,c-d_{xz})}{\max(c,0.001)}

Here b,w,cb,w,c are the required boundary clearance, aisle width, and equipment clearance. The boundary shortfall is gbg_b, the aisle intrusion depth is owo_w, and the equipment shortfall is max⁡(0,c−dxz)\max(0,c-d_{xz}), where dxzd_{xz} is the horizontal clearance between bounding boxes. Workspace overflow and environmental penetration have deficits gout/(1 m)g_{\mathrm{out}}/(1\,\mathrm{m}) and penv/(1 m)p_{\mathrm{env}}/(1\,\mathrm{m}), where penvp_{\mathrm{env}} is the minimum 3D penetration depth involving an environmental obstacle. A 3D collision scores 1 if present and 0 otherwise. The parent/anchor deficit is 1 for an incorrect parent–child relation or anchor and 0 otherwise. Attachment checks apply relation, containment, and contact tolerances.

(3) Functional scores. For horizontal distance dd in metres, the evaluator uses the piecewise-linear function

\begin{cases} 0, & d\le a\\ \text{or}\\ d\ge e, \\\\ (d-a)/(b-a), & a<d<b, \\\\ 1, & b\le d\le c, \\\\ (e-d)/(e-c), & c<d<e. \end{cases}$$ The four material-handling scores are $$S_{\mathrm{process}} = 100T(d_p;1.2,1.5,2.0,3.0)$$ $$S_{\mathrm{tool}} = 100T(d_t;1.2,1.5,1.9,2.4)$$ $$S_{\mathrm{rack}} = 100T(d_r;0,2.0,3.0,4.0)$$ $$S_{\mathrm{container}} = 100T(d_c;0,1.0,2.0,3.0)$$ Here $d_p$ measures the distance from the combined bounding region of the kit-tray station and automated guided vehicle (AGV) to the assembly table. The distance $d_t$ is measured between the tool cart and the kit-tray station. The distances $d_r$ and $d_c$ run from the rack and container, respectively, to the assembly table. For the desk ($D$) and sign ($S$), indices 1 and 2 refer to distances to the AGV and robot proxy, respectively. $$S_{\mathrm{desk}} = 50\sum_{k=1}^{2}\operatorname{clip}\big((d_{Dk}-1.5)/1.5,0,1\big)$$ $$S_{\mathrm{sign}} = 50\sum_{k=1}^{2}T(d_{Sk};0,1.5,3.0,4.5)$$ For each task, Functional Layout is the sum of these six scores divided by six. Invalid or unscorable tasks receive zero and remain in the denominator (Eq. 14). (4) **Output adapters.** Depending on the model interface, we read a structured JSON program directly, extract a complete JSON program from a text response, or convert a native scene representation into the common program format. All outputs undergo the same schema checks, execution, and final validation. Non-executable programs remain failures and bypass repair. We will include the search limits, normalization rules, scoring functions, and adapter descriptions in the revision. We sincerely thank you for your **constructive feedback and valuable suggestions**, which have helped us improve the clarity and rigor of our work.
Reviewer_ueKDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VIGOR addresses the gap between executable scene programs and geometrically feasible industrial layouts. It combines supervised fine-tuning with monotonic local projection and conflict-graph-based joint repair, checking the serialized final scene against structural and geometric constraints. The accompanying benchmark varies clearance, aisle, boundary, and orientation requirements within a fixed scene schema, retaining 1,749 configurations solved by a reference constructor. On the compositional split, the complete method reaches 96.86% feasibility. Ablations show complementary benefits from supervised learning, local repair, and joint repair.

Critical strengths
  1. The paper evaluates geometric feasibility rather than just successful execution. Checking saved and reloaded scenes against explicit constraints, including boundary and clearance requirements, makes the benchmark practically meaningful.

  2. Combining SFT with local and joint repair raises feasibility from 14.57% to 96.86%. Both repair stages also help individually, supporting their complementary roles in the complete pipeline.

Critical weaknesses
  1. The benchmark fixes the scene structure and varies a parameter grid already solved by the reference constructor. Without constructor or retrieval-plus-repair baselines, the added value of learned generation under the same final validator remains unclear.

  2. Repair-search settings and runtime costs are insufficiently specified. Candidate generation, beam width, and search limits can materially affect feasibility, so these details matter for reproduction and fair comparison.

Questions for authors
  1. How do constructor and retrieval-plus-repair baselines compare in feasibility and runtime? Please use the same final validator and, where applicable, the same repair budget so that the comparison isolates the generation method.

  2. What are the joint-repair move set, beam width, and search budget? Please also clarify the termination rules and report the computational cost of repair, since these determine how easily the result can be reproduced.

  3. Does VIGOR still outperform a stronger generator using the same repair procedure? This would help distinguish improvements from the learned generator from those provided by the downstream repair stages.

4 comments

We thank you for highlighting the need to clarify the added value of learned generation, the reproducibility and cost of repair search, and comparisons with stronger generators.

Q1. Learned generation compared with reference construction and retrieval (W1&Q1).

A1: The constructor solves all 350 held-out tasks, while VIGOR achieves the highest mean Functional Layout (72.39). We evaluated the deterministic reference constructor, nearest-neighbor retrieval, and retrieval followed by local and joint repair (L/J) on the same tasks with the same final validator.

MethodComplete FeasibilityFunctional LayoutAvg. Runtime
Constructor100.00%70.9364.440 s
Retrieval-only16.29%70.260.054 s
Retrieval+L/J88.29%71.481.079 s
VIGOR (Ours)96.86%72.3990.839 s

Constructor’s 100% feasibility follows from the benchmark design. A condition enters the dataset only if the construction pipeline produces a valid layout (Appendix C). Nearest-neighbor retrieval reuses layouts from a prebuilt library and takes 0.054 s per task. Against Retrieval+L/J, VIGOR gains 8.57 percentage points in Complete Feasibility and 0.91 points in Functional Layout, with gains of 15.98 on Process Chain and 7.91 on Tool/Kitting. VIGOR takes 90.839 s per task, compared with 1.079 s for repaired retrieval.

Q2. Joint-repair search settings and computational cost (W2&Q2).

A2: Joint repair rebuilds the conflict graph after verifying each candidate layout. It proposes single-device translations, two-device separations, coordinated displacements, and constrained rotations. The beam retains up to 16 states under the lexicographic objective in Eq. (7), with at most 60 candidates per expansion and a maximum depth of 12. Search stops as soon as it finds a feasible layout, exhausts the candidates, or reaches the depth limit, 2,000 validator calls, or 60 s.

Local and joint repair together average 4.132 s per task, excluding generation. For scenes that reach joint search, J averages 4.292 s, with a median of 319.5 validator calls. The 90.839 s reported in A1 includes initial generation. We will add these settings and separate timings to Section 3.4.

Q3. Contributions of the initial generator and subsequent repair (W1&Q3).

A3: With L/J held fixed, fine-tuning Qwen3-VL-8B-Instruct raises Complete Feasibility from 0.00% to 96.86%. Fine-tuning alone achieves 14.57% (Table 3). We also replaced the initial generator with each model below while holding the test tasks, L/J search budget, and final validator fixed.

MethodComplete FeasibilityFunctional QOverall
DeepSeek-V4.1-Flash + L/J99.14%40.1784.82
GPT-5.5 + L/J62.86%36.6971.85
Ministral-3-8B + L/J0.00%29.1229.76
VIGOR96.86%68.6291.20

DeepSeek-V4.1-Flash + L/J leads on feasibility by 2.28 percentage points. VIGOR leads by 28.45 points on Functional Q and 6.38 on Overall under the same repair procedure. Functional Q averages the full-test, bottom-quartile, interpolation, compositional, and tight-condition Functional Layout scores (Eq. 15).

We would be grateful if you would reconsider your evaluation in light of the constructor comparison and the controlled generator ablations.


As the discussion phase is approaching its end, we kindly request the reviewer to let us know if the above clarifications and the previously added experiments have addressed the remaining questions. We would be happy to address any additional points the reviewer may have during the remaining time of the discussion phase. We thank the reviewer for engaging with us in the discussion.

Dear Reviewer,

As the discussion phase is approaching its end, we would like to kindly follow up on our previous responses. We have carefully considered your concerns regarding the value of learned generation and the reproducibility of our repair method. If possible, we would sincerely appreciate your feedback on whether our clarifications and additional experiments have adequately addressed your main concerns. Should any questions remain, we would be happy to provide further clarification during the remaining discussion period.

Thank you again for your time and valuable feedback on our work.

Dear Reviewer,

As the discussion period draws to a close, we would like to sincerely thank you again for your time and consideration. If you have had an opportunity to review our clarifications and additional experiments, we would be grateful for any final feedback you might wish to share. We would be happy to address any remaining concerns.

We greatly appreciate your time and thoughtful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.