{"id":"3297a397-a56a-4654-89a9-dfcde28ebdae","arxiv_id":"2501.13337","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GMFoO runs Bayesian optimization on multiple GAN latent spaces simultaneously, using correlated spaces and multi-fidelity knowledge transfer to improve sample efficiency for expensive structured optimization.","lead":"This paper proposes a Bayesian optimization method that searches over several compressed latent versions of a design space at once, instead of one, to speed up expensive design problems like airfoil shape optimization. The method (GMFoO) shares information between a high-dimensional and low-dimensional latent space, and the authors report better designs within limited evaluations on airfoil, corbel, and MNIST digit-area tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GMFoO's knowledge-transfer mechanism rests on an unvalidated correlation between low- and high-dimensional latent spaces; the only reported evidence is a single r=0.8 scatterplot, so the 'always achieves the best solutions' claim is not robustly established.","rationale":"The reader identifies the correlation between the low-dimensional latent space C and the high-dimensional latent space Z as the weakest assumption, and I agree. Both of GMFoO's information-exchange strategies—narrowed exploitation and multi-fidelity co-kriging—require that C is a faithful low-rank summary of Z, not merely that the two spaces are 'related' in some loose sense. The paper's support is a single scatterplot with r=0.8 for one airfoil case and a parallel-coordinate plot for the same case; no correlation is measured for the subsonic airfoil, corbel, or MNIST tasks, and no negative-transfer analysis is provided. Because the central empirical claim is that GMFoO 'always achieves the best solutions with even faster convergence rates' (Section 4.3), a failure of this premise on any tested problem would directly contradict the 'always' claim. The reader's CONDITIONAL verdict is appropriate: the approach is plausible and the experiments are encouraging, but the key mechanism is under-validated. The concrete test proposed above would either confirm the correlation premise across all problems or reveal that the reported r=0.8 does not generalize to the optimization-relevant regions, in which case the claimed advantage would need to be attributed to other algorithmic components (e.g., the extra exploitation evaluations) rather than positive transfer. No change to the verdict is needed; the conditionality already reflects this concern.","tokens_in":17739,"tokens_out":12855,"duration_ms":116985,"concrete_test":"Using the authors' MFoO-GAN and simulation setup, record all (z, y(z)) pairs acquired during the reported GMFoO runs, compute the corresponding c' via the discriminator inverse inference, and compute the Pearson and Spearman correlations between y(z) and y(c') for each of the four test problems (low-speed airfoil, subsonic airfoil, corbel, MNIST). If the Spearman correlation is below approximately 0.7 on any problem, or if the r=0.8 value from Section 4.6.1 is not reproduced on the optimization-acquired samples, then the 'well correlated' premise underlying Eqs. (16)-(19) and the narrowing step is not supported, and the claimed advantage of GMFoO is not explained by the proposed transfer mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of GMFoO is that simultaneous optimization in a high-dimensional latent space Z and a low-dimensional latent space C, with continuous information exchange, yields better solutions with faster convergence (Sections 3.2, 4.3, 4.6). Both exchange strategies presuppose that C is a faithful low-dimensional summary of Z: (1) the 'enhanced local exploitation' (Section 3.2, Fig. 4) narrows the Z search to a box around [c_min, 0, ..., 0] with radius Δ, which is only safe if the global optimum of Z lies near this low-dimensional slice; (2) the multi-fidelity GP (Eqs. 16-19) treats {c', y(z)} as low-fidelity data for C, which is only valid if y(z) ≈ y(c') over the sampled region. The paper's only direct evidence is a single scatterplot for the low-speed airfoil (Section 4.6.1) with Pearson r=0.8 (R²≈0.64, leaving substantial unexplained variance), plus a parallel-coordinate plot of optimal c1-c3 values. No correlation is reported for the subsonic airfoil, corbel, or MNIST problems, and there is no failure analysis or bound on negative transfer. If the correlation is weaker in the regions actually visited during optimization, the co-kriging surrogate can be biased and the narrowed search can exclude the true optimum, causing GMFoO to underperform the single-space GMO-High baseline. Thus the experimental demonstration of the central claim rests on an unvalidated premise that is directly load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GMFoO (generative multi-form optimization), a framework that trains a GAN-like generative model (MFoO-GAN) to produce a high-dimensional latent space Z and a low-dimensional latent space C that are positively correlated via an InfoGAN-style mutual information regularizer. Bayesian optimization is then run simultaneously in both latent spaces, with two knowledge-transfer mechanisms: a multi-fidelity Gaussian process that treats samples transformed between Z and C as low-fidelity data, and an 'enhanced local exploitation' step that narrows the Z search region around the current best point of C. Experiments on low-speed and subsonic airfoil design, decorative corbel design, and an MNIST area maximization problem compare GMFoO against GMO-High, GMO-Low, GMO-NashEGO, SVD-BO, GMO-CMAES, and GMO-IKEA, with ablations for the size parameter Δ and the low-dimensional latent dimension dL.","tokens_in":18111,"tokens_out":7163,"duration_ms":64025,"significance":"The paper addresses a genuine and practically relevant issue: the choice of latent dimension in generative-model-based optimization involves a trade-off between solution accuracy and convergence rate. If the empirical claims hold, GMFoO offers a useful way to hedge this choice by optimizing over multiple correlated latent spaces simultaneously, and the multi-fidelity treatment of transformed samples is an interesting idea. The manuscript includes ablations for the main algorithmic components and a correlation analysis between the latent spaces. However, the strength of the empirical claims is not yet matched by the evidence: some statements overreach the data, the comparison protocol is under-specified, and the central correlation premise is validated only for one of the four test problems.","major_comments":[{"comment":"The claim in Section 4.3 that 'GMFoO ... always achieves the best solutions with even faster convergence rates' is contradicted later in the same section, where the authors state that for the low-speed airfoil GMO-IKEA achieves final solutions slightly better than those of GMFoO (Fig. 8(b)). Please moderate this claim and analyze the conditions under which GMFoO is and is not the best method.","section":"4.3, Fig. 8(b)"},{"comment":"GMFoO evaluates three points per main loop iteration (one in C, one in the full Z, and one in the narrowed Z, per Algorithm 3 lines 6 and 9), while the baselines evaluate one point per iteration. If the convergence plots use iterations on the x-axis, GMFoO receives a 3x larger budget per iteration; if they use function evaluations, the number of evaluations per iteration should be stated. Please report results as a function of the total number of expensive function evaluations and describe the DoE sizes exactly.","section":"4.2, Algorithm 3, Figs. 8-9"},{"comment":"The two knowledge-transfer mechanisms in Section 3.2 presuppose that the low-dimensional latent space C is well correlated with Z, but the only direct evidence is the single Pearson coefficient r=0.8 for the low-speed airfoil. No correlation analysis is reported for the subsonic airfoil, corbel, or MNIST cases, and the fitted multi-fidelity GP correlation coefficients ρ_ij are not reported. Please add per-problem correlation and negative-transfer diagnostics, or explain why the single case is sufficient.","section":"4.6.1, Fig. 14"},{"comment":"The parameters Δ and dL are selected per problem after sensitivity analyses, and the main comparisons use the best configuration (e.g., Δ=0.15, dL=4). No statistical significance tests are applied to the 10-run comparisons, so the reported differences among algorithms may not be significant. Please report significance tests (e.g., Wilcoxon signed-rank or rank-sum tests) and state whether the sensitivity-analysis results are used to pick parameters before or after the main comparison.","section":"4.2, 4.6.2-4.6.3"}],"minor_comments":[{"comment":"The covariance matrices written with ρ12 and ρ21 are only symmetric if ρ12 = ρ21; this constraint should be stated explicitly. Also, the shorthand K(C') and K(Z') should be defined as K(C',C') and K(Z',Z').","section":"Eqs. (16) and (18)"},{"comment":"The derivation treats H(c) as constant when moving from Eq. (14) to Eq. (15); this is only valid if the marginal distribution of c is fixed during generator training, which should be clarified.","section":"Eqs. (14)-(15)"},{"comment":"It is unclear whether the point selected in the narrowed high-dimensional space is included in the training set for the multi-fidelity GP of Z; please specify the update step.","section":"Algorithm 3, line 9"},{"comment":"The convergence-history plots are missing explicit x-axis labels; please state whether the horizontal axis is the number of iterations or the number of expensive function evaluations.","section":"Figs. 8 and 9"},{"comment":"The sentence 'The number of initial training samples for the standard BO and IKEA are set as 11 times of the dimension of the low-dimensional latent space of MFoO-GAN (i.e., 11 dL)' is surprising because GMO-High optimizes in a 13- or 23-dimensional space but receives only 33 or 44 initial samples; please clarify why the DoE size is tied to dL rather than the dimension being optimized.","section":"4.2"},{"comment":"In the contribution list, 'we instantiates the proposed GMFoO' should be 'we instantiate the proposed GMFoO'.","section":"1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from releasing code and exact experiment configurations; the current version does not allow a reader to reproduce the results. The paper also relies heavily on earlier work by the same group (e.g., Refs. [8], [29], [43], [44], [60]); this is not a problem per se, but it increases the importance of independent validation of the new components. The fit for a journal such as IEEE TEVC or Structural and Multidisciplinary Optimization is appropriate if the empirical claims are tightened and the comparison protocol is clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper introduces GMFoO, a generative-model-based optimization that runs BO in a high-dimensional and a low-dimensional latent space simultaneously. The two spaces are correlated via InfoGAN-style mutual information regularization; samples are exchanged through a discriminator-based inverse mapping; and the low-dimensional optimum is used to narrow the high-dimensional search. That configuration is genuinely new relative to the cited literature. The framework is clearly explained, the experiments span airfoil design, corbel design, and an MNIST area-maximization problem, and the ablations on the narrowing parameter Delta and the low-dimensional dimension dL show the components contribute. The correlation analysis (Pearson r=0.8, Fig. 14) is a good attempt to check the central transfer premise, and the authors are transparent about the tuning parameters.\n\nThe soft spots are mostly empirical. No code is released, which matters for a method with this many moving parts. Key parameters dL and Delta are tuned per problem, and the differences between GMFoO and the best baseline are often small; no significance tests are reported. On the low-speed airfoil, GMO-IKEA actually ends with a slightly better final solution (Fig. 8b), which undercuts the absolute claim that GMFoO 'always achieves the best solutions.' The load-bearing correlation premise is supported by one scatterplot from one problem; no correlation evidence is given for the subsonic, corbel, or MNIST cases, and negative transfer is not analyzed. The MNIST experiment omits weighted retraining, which is the natural strong baseline for latent-space optimization. These are fixable, but they keep the current version from being a clean accept.\n\nI don't fully agree with the stress-test note that the correlation premise is unvalidated. The r=0.8 plot plus the parallel-coordinate analysis of optimal solutions (Fig. 15) is some evidence in the right direction; the issue is that the evidence is thin, not absent. The paper is honest about its limitations, and the comparisons are held-out, so I don't see circular reasoning.\n\nThis is a serious paper. The idea is worth engaging with, the engineering is careful, and the experiments are mostly consistent. A serious referee should get it, with the expectation that the authors release code, add significance testing or more runs, include weighted retraining on the MNIST task, and either soften the 'always achieves' claim or support it with broader correlation statistics. I would send it to review despite my reservations.","headline":"A genuinely new multi-latent-space GMO scheme with InfoGAN-style correlation and co-kriging transfer; the core idea is plausible and the experiments mostly support it, but the 'always best' claim overreaches the evidence.","tokens_in":18643,"tokens_out":1635,"would_cite":true,"duration_ms":15939,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GMFoO resolves the latent-dimension accuracy-versus-speed trade-off in generative optimization by running Bayesian optimization on correlated high- and low-dimensional latent spaces simultaneously.","keywords":["generative model based optimization","multi-form optimization","Bayesian optimization","latent space","multi-fidelity Gaussian process","knowledge transfer","airfoil design","InfoGAN"],"falsifier":"Train MFoO-GAN on a design dataset, then compare GMFoO against GMO-High on a problem where the learned low-dimensional space is artificially decorrelated from Z, for instance by shuffling the low-dimensional codes relative to their high-dimensional partners. If GMFoO still wins, its success does not depend on the correlation assumption; if it loses, the assumption is confirmed as load-bearing. A cheaper check is to compute the Pearson correlation between y(z) and y(c') on held-out samples and test whether GMFoO's advantage disappears below some r threshold.","tokens_in":17510,"feed_emoji":"📐","tokens_out":5685,"duration_ms":48803,"temperature":0.7,"pith_summary":"Generative model based optimization (GMO) maps a complex structured design space into a continuous latent space, but choosing the latent dimension forces a trade-off: low dimensions converge fast yet cap solution quality, high dimensions are accurate but slow. This paper claims that optimizing over two correlated latent spaces at once breaks that trade-off. The proposed method, GMFoO, trains a GAN whose low-dimensional latent code is regularized to correlate with its high-dimensional latent vector, then runs Bayesian optimization in both spaces, using the low-space optimum to narrow the high-space search and treating cross-projected samples as low-fidelity data in a multi-fidelity Gaussian process. On airfoil, decorative corbel, and MNIST-area test problems, the paper reports that GMFoO reaches better designs faster than single-space GMO, randomly-subspaced NashEGO, SVD-BO, and evolutionary baselines within the same sample budget. If the claim holds, GMFoO offers a practical hedge against latent-dimension misspecification in expensive black-box design.","feed_headline":"Multiple latent spaces beat one in generative design optimization","feed_subtitle":"Bayesian optimization over correlated high- and low-dim spaces finds better airfoils and corbels within a fixed budget.","key_machinery":"The load-bearing object is MFoO-GAN, a generative adversarial network whose training loss adds the InfoGAN mutual-information term I(c; G(z)) so that the low-dimensional code c captures the major variability of the generated object and is correlated with the high-dimensional vector z = [c, z*]. On top of this, two exchange mechanisms carry the argument: enhanced local exploitation, which narrows the search box in Z around the current best low-space solution c_min with radius Δ, and multi-fidelity GP surrogates, where samples transformed from C into Z (by fixing z*) and from Z into C (by inverse inference through the discriminator) are treated as low-fidelity data and combined with exact expensive samples through co-kriging correlation coefficients ρ_ij. The mutual-information regularization is what makes the transferred samples trustworthy, and the multi-fidelity GPs are what convert that trust into faster convergence.","core_discovery":"The central claim is that the apparent conflict between solution accuracy and convergence rate in GMO can be resolved by multi-form optimization over multiple latent spaces of one generative model, rather than by choosing a single latent dimension. In GMFoO, a GAN with InfoGAN-style mutual information regularization, called MFoO-GAN, produces a high-dimensional space Z and one or more low-dimensional spaces C that are positively correlated; the best point found in C is used to narrow the search region in Z around [c_min, 0], and samples projected between Z and C are fed into two multi-fidelity GPs as low-fidelity data. The paper's experiments on low-speed and subsonic airfoil design, corbel design, and binary-image area maximization show GMFoO achieving the best final solutions with faster convergence than GMO-High, GMO-Low, GMO-NashEGO, SVD-BO, GMO-CMAES, and GMO-IKEA.","pith_inferences":["The correlation requirement suggests a cheap diagnostic: measure the Pearson correlation between y(z) and y(c') before trusting transfer; the paper reports one case with r=0.8, so users could set a threshold below which low-fidelity injection is switched off.","The same multi-space hedge could be built on other generative models, such as VAEs or diffusion models, by imposing analogous mutual-information or reconstruction constraints between latent subspaces; nothing in the argument is specific to GANs.","A natural extension is to adapt Δ and the low-dimensional dimension dL online from estimated correlation and surrogate disagreement, since the paper fixes them by default and shows sensitivity to both.","Because negative transfer is possible when the low-space optimum drifts away from the high-space optimum, adaptive gating or weighting of transferred samples would make the method safer on problems where the generative model's latent spaces are not well correlated."],"forward_implications":["GMFoO's best solution and convergence rate dominate single-space GMO baselines on all three test problems within the stated budgets.","The multi-fidelity component alone, corresponding to Δ=0, outperforms GMO-High, and adding the c_min-guided narrowing improves further, so each transfer strategy contributes independently.","Correlated latent spaces transfer more effectively than randomly generated alternate subspaces, explaining why GMO-NashEGO lags behind GMFoO.","The gains persist when the high-dimensional latent space grows from 13 to 23 dimensions, where BO-based GMFoO overtakes evolutionary GMO-IKEA within budget.","The framework applies to structured design spaces with hundreds of interacting variables, including 384-variable airfoil contours, corbel curves, and 784-bit binary images, rather than only low-dimensional Euclidean test functions."],"supporting_citations":[{"why":"Introduces Bezier-GAN airfoil parameterization and optimization, providing the generative latent mapping that GMO builds on and that GMFoO extends.","marker":"[4]"},{"why":"Demonstrates aerodynamic design optimization with GANs, establishing the two-stage GMO procedure.","marker":"[8]"},{"why":"Provides the transfer-optimization taxonomy that frames multi-form optimization and motivates correlated knowledge transfer.","marker":"[15]"},{"why":"Supports using low-fidelity samples in multi-fidelity surrogates, justifying the co-kriging treatment of transferred samples.","marker":"[44]"},{"why":"Supplies NashEGO, the random-subspace multi-form BO baseline that GMFoO is compared against.","marker":"[47]"},{"why":"Defines expected improvement, the acquisition function used to select new samples in both latent spaces.","marker":"[51]"},{"why":"InfoGAN mutual-information regularization is the mechanism used to correlate the low- and high-dimensional latent spaces.","marker":"[52]"}],"fun_headline_variants":["Two correlated latent spaces improve Bayesian optimization convergence","Generative multi-form BO finds better designs in fewer trials","Multi-latent Bayesian optimization converges faster on hard design tasks","Pairing latent spaces speeds Bayesian design optimization","Better airfoils and corbels via multi-latent Bayesian optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method presupposes that the low-dimensional latent space C is correlated enough with the high-dimensional space Z that c_min lies near Z's optimum and that projected C-samples improve the multi-fidelity surrogate rather than corrupt it; the paper demonstrates one Pearson r=0.8 case and does not characterise when transfer turns negative.","fun_headline_variants_meta":{"raw":{"variants":["Two correlated latent spaces improve Bayesian optimization convergence","Generative multi-form BO finds better designs in fewer trials","Multi-latent Bayesian optimization converges faster on hard design tasks","Pairing latent spaces speeds Bayesian design optimization","Better airfoils and corbels via multi-latent Bayesian optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000389,"raw_usage":{"total_tokens":2051,"prompt_tokens":948,"completion_tokens":1103,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1023}},"tokens_in":564,"tokens_out":1103,"duration_ms":12613,"temperature":1.0,"reasoning_tokens":1023,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:14:46.002255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MFoO-GAN on a design dataset, then compare GMFoO against GMO-High on a problem where the learned low-dimensional space is artificially decorrelated from Z, for instance by shuffling the low-dimensional codes relative to their high-dimensional partners. If GMFoO still wins, its success does not depend on the correlation assumption; if it loses, the assumption is confirmed as load-bearing. A cheaper check is to compute the Pearson correlation between y(z) and y(c') on held-out samples and test whether GMFoO's advantage disappears below some r threshold.","supporting_citations":[{"cited_title":"Aerodynamic design optimization and shape exploration using generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Demonstrates aerodynamic design optimization with GANs, establishing the two-stage GMO procedure."},{"cited_title":"Nash game based efficient global optimization for large-scale design problems,","cited_arxiv_id":null,"evidence_quote":"Supplies NashEGO, the random-subspace multi-form BO baseline that GMFoO is compared against."},{"cited_title":"Infogan: Interpretable representation learning by information maximizing generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"InfoGAN mutual-information regularization is the mechanism used to correlate the low- and high-dimensional latent spaces."}],"review_version":1}