{"id":"9e2a4f31-f9a0-40d5-8cb8-b0fe9227dd6c","arxiv_id":"2505.18188","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A VAE-based two-stage pipeline with latent search and best-of-N sampling generates rectangular patch antenna geometries that approximate target S11 responses, with accuracy improving as search budget grows.","lead":"This paper trains a two-stage generative model that turns a target antenna frequency response into rectangular patch antenna dimensions, then uses search and optimization at test time to improve generated designs. It is a useful test of whether extra inference-time compute can compensate for small training datasets in inverse design.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Figure 3's scaling evidence is based entirely on a surrogate scorer with no demonstrated correlation to openEMS error; the one full-wave check that fails is consistent with surrogate over-ranking.","rationale":"The reader's weakest assumption identifies exactly the load-bearing risk: the scaling evidence is internal to the surrogate. I agree with that judgment. I considered whether the lack of baselines (e.g., best-of-N random search over L, W, p) is the more fundamental issue; adding such baselines would strengthen the paper, but the surrogate-validation gap is more immediate because it undermines the paper's own quantitative figure, and the admitted 5.0 GHz failure is direct evidence that the gap is not hypothetical. The promised code and data make the proposed openEMS check feasible. The paper has independent value as a two-stage generative pipeline, and the failure is acknowledged, but the accuracy claim should remain conditional until surrogate scores are calibrated against full-wave simulation.","tokens_in":9649,"tokens_out":3729,"duration_ms":42439,"concrete_test":"For the 5.0 GHz target (and one easy target such as 2.4 GHz), reproduce the Sec. 6.2 search at budgets 1x1, 5x5, and 10x20 with random latent initialization. Rank each pool by the Surrogate Scorer, then run openEMS on the top 50 designs per budget and compute the masked MSE against y* and the achieved depth at f0. Report Spearman correlation between surrogate score and openEMS error across all candidates, and plot mean true masked error versus budget. If the correlation is weak or if true error does not decrease with budget, the Fig. 3 scaling claim would not support the paper's accuracy conclusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that test-time compute improves design accuracy. The quantitative support is Fig. 3, whose y-axis is 'Average Lowest Surrogate Score' from a neural surrogate trained to approximate openEMS. A best-of-N optimization will lower the score of any chosen scorer as the pool grows, so monotone curves only show that larger pools contain candidates the surrogate prefers; they do not show that the selected designs are electromagnetically closer to the target. No validation is reported linking surrogate score to true |S11| error on generated candidates. The only independent check is Fig. 4, with three targets and two budgets. For the 5.0 GHz target, even the 10x20 search selects a design that fails the d=-10 dB depth requirement (Sec. 6.3), exactly the result expected if the surrogate over-ranks candidates. Because Sec. 6.2's scaling argument and Sec. 6.3's design selection both rely on this unvalidated metric, the accuracy-improvement claim is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage deep learning framework for inverse design of rectangular patch antennas. Stage 1 trains a VAE to encode S11 frequency response curves into a latent space; at test time, a latent-space search finds an in-distribution response matching a user-specified target (a product of Lorentzian notches). Stage 2 trains a conditionally disentangled CVAE that maps a target response and a latent code to antenna geometry (L, W, p). The central claim is that test-time compute - specifically best-of-N sampling of curves and designs and gradient-based optimization of the conditional latent code - improves the accuracy of generated designs without additional training data, while also enabling auxiliary objectives such as manufacturability. The experiments report scaling curves (Figure 3) using a neural surrogate scorer, and show full-wave openEMS simulations for three target responses at two search budgets (Figure 4). The paper concludes that the approach generalizes to more complex design spaces and provides public code and data.","tokens_in":9878,"tokens_out":3581,"duration_ms":39204,"significance":"If established, the claim that test-time search and optimization can substantially improve inverse-design accuracy with limited training data would be practically valuable, given the cost of EM simulation datasets. The paper has notable strengths: it uses an external full-wave solver (openEMS) for final validation, it explicitly acknowledges a failure case (Sec. 6.3), and it releases the dataset, code, and trained models. The two-stage architecture (latent-space response search plus conditioned geometry generation) is a sensible way to handle out-of-distribution targets and one-to-many mappings. However, the quantitative evidence for the central scaling claim is largely self-referential: Figure 3 ranks candidates with a surrogate trained on the same simulation data, with no demonstrated correlation to true EM error, and the only independent full-simulation check (Figure 4) includes one failure that contradicts the monotone-improvement narrative. These issues are fixable with additional validation, but they are load-bearing for the paper's main conclusion.","major_comments":[{"comment":"The scaling evidence in Figure 3 is based entirely on the 'Surrogate Scorer', but the score is not defined anywhere in the paper. Section 5.5 mentions that the surrogate uses a beta-NLL loss and uncertainty estimates, but no exact formula for the score or the frequency masking is given. More importantly, the manuscript never validates that the surrogate score correlates with the true openEMS error on generated candidates. Best-of-N sampling will monotonically lower any scorer's value as the pool grows, so the monotone curves in Figure 3 are expected even if the surrogate is a poor proxy for EM performance. To support the claim that test-time compute improves design accuracy, the authors should (a) define the surrogate score explicitly, (b) report its correlation with openEMS error on a held-out set of generated designs, and (c) re-plot Figure 3 using full-simulation error (or a validated proxy) for the selected designs at each budget. Alternatively, a direct-optimization baseline that optimizes (L, W, p) with a surrogate or with openEMS would show whether the improvement is specific to the proposed framework.","section":"Section 6.2, Figure 3"},{"comment":"The full-simulation validation covers only three target curves, and one of them - the 5.0 GHz target with d = -10 dB - fails to meet the depth requirement even with the larger 10x20 search budget. This failure is acknowledged in the text, but it directly undercuts the claim that 'devoting more compute to search ... yields a more accurate result' in all cases. The paper should report quantitative metrics for all three targets at both budgets (e.g., minimum |S11| within the target band, bandwidth achieved, and MSE against the target curve). It should also analyze why the 5.0 GHz target fails: is the target out of the distribution of realizable responses, does the surrogate over-rank the selected design, or is the search budget still insufficient? Without such analysis, the reader cannot assess whether the failure is a fundamental limitation or a fixable artifact.","section":"Section 6.3, Figure 4"},{"comment":"The Oracle Scorer, which runs full EM simulation and computes MSE against the target, is defined in Section 5.5 but never used in any experiment. Since the oracle is the ground truth for design quality, comparing the surrogate-ranked selections with oracle-ranked selections on the same three target curves would directly address whether the surrogate over-ranks designs (as the 5.0 GHz failure suggests). Even if the oracle is computationally expensive, a small-scale comparison (e.g., for the 200 designs in the 10x20 budget) would be feasible and would substantially strengthen the paper's evidence. The authors should add this comparison or explicitly justify why the oracle is not used for validation.","section":"Section 5.5, Oracle Scorer"}],"minor_comments":[{"comment":"The abstract says 'adversarial training process', but the method uses an adversarial predictor (a small network) that attempts to infer the condition from the latent code, not a generative adversarial network. Rephrase to 'adversarially disentangled training' or similar to avoid ambiguity.","section":"Abstract"},{"comment":"The shaded regions in Figure 3 indicate variability across runs, but the number of runs is not reported. State how many independent runs were averaged and whether the same target set was used for each run.","section":"Section 6.2, Figure 3"},{"comment":"For the 5.0 GHz target, the generated patch dimensions change substantially between budgets (L = 15.7 mm for 1x1 vs L = 29.0 mm for 10x20, and W from 25.5 mm to 46.9 mm). The paper does not discuss whether either design is physically reasonable or manufacturable, despite the stated auxiliary objective of manufacturability. A brief comment would be helpful.","section":"Table 1"},{"comment":"The sentence 'In each case, it seems that devoting more compute ... yields a more accurate result' uses the hedge 'seems'. Since the preceding sentence and Figure 4 show only three examples, please either provide quantitative support or temper the claim to match the evidence.","section":"Section 6.3"},{"comment":"The dataset description says the grid is 'augmented' with an algorithm to sample additional triplets inside the convex hull, but no details of this algorithm or the resulting distribution are given. Please clarify how the final 1292 designs are split into training/validation/test sets, if at all.","section":"Section 4"},{"comment":"Reference [9] is incomplete: the conference name and page numbers are missing. Please complete the bibliographic entry.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and addresses a relevant problem, but the main quantitative claim requires stronger validation. The self-referential nature of the surrogate-based evaluation is a serious concern that the authors can address in a revision by adding a surrogate-to-simulation correlation study and reporting full-wave results for more targets. The authors are also advised to position their work against direct optimization baselines, which would clarify the novelty of the two-stage generative approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Beck's paper is a clean proof-of-concept, but the central claim—that test-time compute improves design accuracy—is not yet established. The evidence in Figure 3 is entirely surrogate-based, and the surrogate's correlation with real openEMS error is never shown. The 5.0 GHz failure in Section 6.3 is a red flag that the surrogate can over-rank designs. That said, the pipeline is sensible, the failure is reported honestly, and the promised code/data could be a real resource.\n\nWhat's new: not the individual components—latent search, adversarially disentangled CVAEs, best-of-N, test-time latent optimization all exist. The novelty is the combination for rectangular patch antenna inverse design and the systematic look at how performance scales with test-time compute on a fixed dataset. The dataset of 1292 openEMS simulations on a fixed substrate is concrete, and if the repository is complete, that's valuable for anyone in the area.\n\nSoft spots, in order of severity.\n\nFirst, the quantitative backbone is Figure 3, whose y-axis is 'Average Lowest Surrogate Score.' Best-of-N sampling will lower any scorer's loss as the pool grows, so the monotone curves only tell you the surrogate likes the larger pools. There is no check that lower surrogate score corresponds to lower true EM error on generated candidates—no calibration plot, no openEMS spot-check of the designs selected at each budget. The single full-wave failure at 5.0 GHz is exactly the signature of surrogate over-ranking.\n\nSecond, no baselines. The paper would be much more convincing with a simple comparison: direct gradient-free optimization over the surrogate, or nearest-neighbor lookup in the training set, or analytic design equations plus local fine-tuning. Without that, the added value of the generative pipeline is unclear.\n\nThird, the full-wave validation is three targets, one of which misses its depth requirement. That's too few to support the generality that the conclusion claims. At minimum, the scaling curves in Figure 3 should be validated against openEMS for a handful of selected designs per budget.\n\nNone of this is fatal to the core idea. The architecture is well-motivated, the problem is clearly formulated, and the authors openly acknowledge the failure case. The softness is in the evaluation, not the reasoning. If I worked on RF inverse design, I'd want to see the surrogate correlation and baselines before using the method, and I'd want a referee to require them. The paper deserves peer review, but it needs revision before the accuracy claim can be taken as established.","headline":"A plausible two-stage generative pipeline for patch antenna inverse design whose accuracy claim rests on an unvalidated surrogate scorer; the one full-wave failure suggests the concern is real.","tokens_in":10358,"tokens_out":3671,"would_cite":false,"duration_ms":41181,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Spending more inference-time compute improves generated rectangular patch antenna designs, without retraining.","keywords":["inverse design","rectangular patch antenna","test-time compute","variational autoencoder","conditional generation","adversarial disentanglement","S11 reflection coefficient","electromagnetic simulation"],"falsifier":"Run full electromagnetic simulation over the complete candidate pools from the test-time scaling experiments and compare the design selected by the lowest surrogate score with the design having the lowest true simulated error for each target. If the surrogate-selected designs do not also achieve the lowest simulated error, or if the accuracy gap between small and large search budgets disappears under true simulation, the central claim is refuted.","tokens_in":9473,"feed_emoji":"📡","tokens_out":9016,"duration_ms":87711,"temperature":0.7,"pith_summary":"This paper tries to establish that for coaxial-fed rectangular patch antennas, design accuracy can be bought at inference time rather than with more training data. The proposed pipeline first trains a variational autoencoder on measured reflection-coefficient (S11) curves, then trains an adversarially conditioned generative model that maps a target response curve to antenna dimensions (patch length, width, feed position). At test time the paper searches the response-curve latent space, samples many candidate designs, and gradient-optimizes the design latent code against geometric constraints. Across three target specifications, simulated S11 curves improve as the number of sampled curves and designs per curve grows, and the paper argues this behavior parallels test-time compute scaling in other domains.","feed_headline":"More test-time search yields accurate antenna designs","feed_subtitle":"A two-stage generative model gets better as it samples more response curves and designs, with no new training data.","key_machinery":"The load-bearing mechanism is the two-stage latent pipeline: a response variational autoencoder that maps full S11 curves into a low-dimensional latent space, followed by an adversarially disentangled conditional variational autoencoder that maps antenna dimensions (L,W,p) into a separate latent code and decodes geometries from that code together with the response curve. The adversarial predictor is what makes the second stage controllable: it tries to recover the response curve from the design latent code, so the encoder is trained to strip response information out of that code and the decoder must use the supplied curve to reconstruct the geometry. At test time the same latents are treated as search spaces: gradient steps move the response latent toward in-distribution curves that match the masked target, and move the design latent to satisfy geometric constraints without changing the response.","core_discovery":"The central claim is that a two-stage generative architecture becomes more accurate as more compute is spent during inference, without retraining. Stage 1 compresses 1000-point S11 response curves into a 64-dimensional latent space with a beta-weighted variational autoencoder. Stage 2 trains a conditional variational autoencoder whose decoder maps a design latent code together with a response curve to (L,W,p); an adversarial predictor tries to infer the response curve from the design latent code, and the encoder is penalized for making that possible. This forces the decoder to rely on the explicit conditioning input. At test time, gradient search over the response latent finds in-distribution curves matching the masked target bands, best-of-N sampling over both latent spaces grows the candidate pool, and gradient optimization of the design latent against geometric penalties improves manufacturability while preserving the response. The paper demonstrates on three targets that larger search budgets produce simulated responses closer to the target, with the 5.0 GHz target still missing its depth requirement even after extensive search.","pith_inferences":["The same two-stage latent-search recipe should transfer to other inverse-design problems with expensive forward simulators, such as optics, acoustics, or materials, whenever a learnable manifold of feasible responses can be searched.","The 5.0 GHz failure suggests the response autoencoder's latent manifold may lack curves with enough depth at that frequency and bandwidth; adding training data near that region is a concrete test of whether search is limited by the manifold's coverage.","An ablation that trains the conditional generator without the Stage 1 latent search, conditioning directly on the idealized target curve, would isolate how much of the accuracy gain comes from the response search versus the controllable decoder."],"forward_implications":["Increasing the number of sampled response curves from 1 to 20 lowers the average lowest surrogate score for both random and nearest-neighbor latent initialization.","Increasing the number of sampled designs per curve also lowers the lowest surrogate score, and optimizing the design latent against geometric penalties keeps that accuracy while yielding more manufacturable geometries.","Because the pipeline is probabilistic, multiple distinct geometries can satisfy the same response target, and best-of-N sampling exploits that one-to-many structure.","The test-time gains do not require expanding the training dataset or retraining the models, so the method turns inference compute directly into design accuracy."],"supporting_citations":[{"why":"Supplies the variational autoencoder formulation used in Stage 1 to learn a low-dimensional representation of S11 response curves.","marker":"[13]"},{"why":"Supplies the conditional variational autoencoder formulation used in Stage 2 to generate antenna geometries conditioned on a response curve.","marker":"[22]"},{"why":"Supplies the adversarial disentanglement technique that prevents the design latent code from leaking response information, making the conditioning path controllable.","marker":"[7]"},{"why":"Supplies the beta-weighted variational objective that balances reconstruction and latent regularization in the response autoencoder.","marker":"[10]"},{"why":"Provides the electromagnetic simulation used to generate the training dataset and to simulate the final generated designs.","marker":"[14]"},{"why":"Supplies the heteroscedastic loss used to train the surrogate scorer that ranks candidate designs in the test-time scaling experiments.","marker":"[19]"},{"why":"Provides the analytic resonance formula used as a reference when comparing simulated responses against targets.","marker":"[9]"}],"fun_headline_variants":["Test-time search sharpens generative antenna inverse design","More inference compute yields more accurate antenna designs","Two-stage model gains accuracy via test-time search","No retraining: test-time search boosts antenna design accuracy","Generative antenna design improved by test-time optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the neural surrogate scorer ranks candidate designs the same way full electromagnetic simulation would; if the surrogate over-ranks designs that fail in simulation, the reported test-time gains could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Test-time search sharpens generative antenna inverse design","More inference compute yields more accurate antenna designs","Two-stage model gains accuracy via test-time search","No retraining: test-time search boosts antenna design accuracy","Generative antenna design improved by test-time optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1371,"prompt_tokens":817,"completion_tokens":554,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":482}},"tokens_in":433,"tokens_out":554,"duration_ms":5966,"temperature":1.0,"reasoning_tokens":482,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:29:37.169327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run full electromagnetic simulation over the complete candidate pools from the test-time scaling experiments and compare the design selected by the lowest surrogate score with the design having the lowest true simulated error for each target. If the surrogate-selected designs do not also achieve the lowest simulated error, or if the accuracy gap between small and large search budgets disappears under true simulation, the central claim is refuted.","supporting_citations":[{"cited_title":"Learning structured output representation using deep conditional generative models","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional variational autoencoder formulation used in Stage 2 to generate antenna geometries conditioned on a response curve."},{"cited_title":"Gansynth: Adversarial neural audio synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial disentanglement technique that prevents the design latent code from leaking response information, making the conditioning path controllable."},{"cited_title":"In ICLR, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the beta-weighted variational objective that balances reconstruction and latent regularization in the response autoencoder."},{"cited_title":"openems - open electromagnetic field solver, accessed 2024","cited_arxiv_id":null,"evidence_quote":"Provides the electromagnetic simulation used to generate the training dataset and to simulate the final generated designs."},{"cited_title":"Equations for microstrip circuit design","cited_arxiv_id":null,"evidence_quote":"Provides the analytic resonance formula used as a reference when comparing simulated responses against targets."}],"review_version":1}