{"id":"929d73e5-ef25-491b-80f0-ac62388a2d55","arxiv_id":"2501.18344","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Transfer learning of Gaussian process surrogates is improved by fitting nonlinear beta-CDF input warping plus affine transformations on small target datasets.","lead":"This paper shows that a trained surrogate model can be adapted to a new task by learning a nonlinear input warp (beta CDF) plus an affine transformation from a few target samples. The method beats retraining from scratch when target data are scarce, on BBOB benchmarks and an automotive braking problem.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BBOB validation is circular: targets are generated inside the method's own beta-CDF+rotation family, so the central low-data advantage is untested under misspecification; the ABS benchmark shows the family can fail (instance3), yet the abstract claims universal gains.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the BBOB targets are generated inside the method's own transformation family, so those experiments assume the premise rather than test it, and the real-world ABS benchmark shows the family can fail. My stress-test confirms this and sharpens it: the only out-of-family benchmark provides mixed evidence, with instance3 being a clear counterexample where scratch training is best. This directly undercuts the abstract's universal phrasing. The paper's own discussion in Section 5.3 is admirably honest about the BBOB circularity and the ABS failure, which is why I do not see the issue as fatal to the paper's contribution; rather, it requires a scoped claim and a misspecification experiment. Since the reader already recommended CONDITIONAL acceptance with this caveat, and my analysis supports that condition rather than rejecting the paper, the verdict should remain as the reader set it. The gradient derivations and the Riemannian optimization appear correct, and the ablation study (Fig. 5) does show the full method can beat affine-only in the in-family setting, so there is a real method here; the weakness is in the generality of the empirical support, not in the internal mathematics. A concrete out-of-family test would settle whether the claimed advantage extends beyond the exact generative family.","tokens_in":36959,"tokens_out":3263,"duration_ms":35560,"concrete_test":"Re-run the BBOB transfer experiments using the BBOB suite's own native instance transformations as the target generator, rather than the synthetic beta-CDF warp. For example, use the first instance as the source and instance 2 (or another instance) as the target, with no additional beta-CDF construction. BBOB's internal instance transformations are not generated from the assumed W∘phi(.)+v family, so this is a controlled misspecification test. Compare the transferred GPR against a GPR trained from scratch on transfer sets of size 20, 40, and 80 for all 24 functions using the same SMAPE metric and significance tests. If the transferred model is not consistently better than scratch under this out-of-family condition, the conclusion must be scoped to transformations inside the assumed family. If it remains better, the misspecification concern is substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that re-parameterizing a source surrogate as f_hat_S(W phi(x)+v), with phi a coordinatewise beta CDF, and fitting (W,v,alpha,beta) on a small transfer set yields a surrogate more accurate than one trained from scratch in low-data settings. The main support is the BBOB experiment (Section 4), but there the target functions are generated by exactly the assumed class: 'we construct the target f_T by applying a beta CDF transformation, followed by random rotation and translation transformations, to the base function.' Thus the BBOB results only show that the optimizer can recover parameters inside the correct model family, not that the family is adequate for real transfer problems. The paper explicitly acknowledges this in Section 5.3: 'In the BBOB problem suite, our target functions are explicitly designed so that a perfect transformation exists.' The only out-of-family evidence is the ABS automotive benchmark, and Fig. 23 shows that on instance3 the scratch-trained model is the best performer and the transferred model fails, while on several other instance pairs the full method is not consistently superior to training from scratch. Therefore the abstract's statement that 'the transferred surrogate significantly outperforms both the original surrogate and the one built from scratch' is not supported when the true transformation is misspecified, which is the typical real-world case. The paper's own conclusion is more cautious, but the abstract and Section 1 overstate the scope of the result. This is load-bearing because the method's practical value depends on robust behavior when the assumed beta-CDF-plus-rotation structure is only approximate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a transfer learning method for surrogate models in which a pre-trained source surrogate f_hat_S is re-parameterized as f_hat_S(W phi(x) + v), where phi is a coordinatewise beta CDF input warping, W is a rotation in SO(d), and v is a translation. The parameters are fit by minimizing mean squared error on a small transfer set from the target function. For differentiable surrogates such as Gaussian process regression, the authors derive gradients and use Riemannian gradient descent on SO(d); for non-differentiable surrogates they outline a CMA-ES approach on the Lie algebra of SO(d). Experiments compare the transferred GPR with the original GPR and with a GPR trained from scratch on BBOB synthetic transfer problems (in 2D, 5D, and 10D, with four warp geometries and several sample sizes) and on an automotive ABS benchmark, with additional ablations for beta-CDF-only and affine-only transfer.","tokens_in":37341,"tokens_out":6036,"duration_ms":60798,"significance":"The method is a natural and simple extension of affine-only transfer learning to nonlinear input warpings, and the gradient derivation in Eqs. (5)-(14) is sound. The paper provides broad experimental coverage: 24 BBOB functions, multiple dimensions, several transfer sample sizes, four beta-CDF shape regimes, ablations, and an independent real-world automotive benchmark. The authors also make code and supplementary material available via Zenodo, and the 'in-domain' ablation addresses an important boundary-effect issue. If the central claims were fully supported, this would be a useful GECCO contribution. However, the BBOB validation is largely self-referential because the targets are constructed inside the method's own model family, and the real-world ABS results are mixed; the abstract's unqualified claim that the transferred surrogate 'significantly outperforms' both baselines is stronger than the evidence supports.","major_comments":[{"comment":"The BBOB experiments do not test the adequacy of the assumed transformation family, because the targets are generated inside that family: the text states 'we construct the target f_T by applying a beta CDF transformation, followed by random rotation and translation transformations, to the base function.' The paper itself acknowledges in §5.3 that 'In the BBOB problem suite, our target functions are explicitly designed so that a perfect transformation exists.' Consequently, the BBOB results primarily demonstrate parameter recovery within the correct model class, not robustness to misspecification. This is a load-bearing gap for the abstract's claim of general effectiveness. I recommend adding out-of-family experiments, for example targets generated with a different warping (tanh, spline, Kumaraswamy) or with BBOB's own instance transformations, and/or substantially qualifying the abstract and Section 1 claims.","section":"§4 and §5.3"},{"comment":"The beta CDF in Eq. (2) is defined as an integral from 0 to x_i, which is only valid on the support [0,1] of the beta distribution, yet the experiments sample inputs in [-5,5]^d. The paper never states how inputs are normalized into [0,1] before applying phi, nor how the subsequent rotation and translation map the warped coordinates back to the source domain. This is not a minor formalism issue: without the normalization and domain-mapping details, the experiments and the gradient formulas in Eqs. (5)-(14) are not fully reproducible. The 'in-domain' ablation in §5.2 is a partial treatment of boundary effects, but the base normalization should be specified explicitly.","section":"§3, Eq. (1)-(2), and §4"},{"comment":"The real-world ABS results do not support the abstract's universal claim. Fig. 23 shows that for transfers involving instance3, the scratch-trained GPR is the best performer and the transferred model fails, and the text in §5.3 states that 'the scratch-trained model remains the top performer among all GPR variants' and that 'there are specific scenarios, such as transferring related to problem instance3, where the transfer learning approach fails.' The conclusion section is appropriately hedged, but the abstract and Section 1 claim that the transferred model 'significantly outperforms both the original surrogate and the one built from scratch' without these caveats. The claims should be made conditional on the transformation family being approximately correct and on the data-scarce regime.","section":"§5.3 and Fig. 23"}],"minor_comments":[{"comment":"The projection formula for the Riemannian gradient is ambiguous as typeset; it should read P(M) = W (W^T M - M^T W) / 2. Please clarify the parentheses.","section":"§3.1, Eq. (15)"},{"comment":"The remark that the beta CDF 'preserves the convexity of the surrogate' is not correct in general: monotonicity alone does not preserve convexity under composition, since a concave coordinate warping composed with a convex function need not remain convex. Please revise or remove this claim.","section":"§3, Remark (1)"},{"comment":"The extension to non-differentiable surrogates via CMA-ES is described but never evaluated in the experiments. If this is intended as a contribution, add at least one experiment or explicitly label it as future work.","section":"§3.2"},{"comment":"The sample-size description is inconsistent: the text mentions 'a larger dataset of 80 points for the 10-dimensional problems,' but Table 3 and Fig. 11 report 40, 80, and 400 samples for 10D. Please reconcile the text with the actual experimental settings.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the mismatch between the abstract and the evidence: the BBOB validation is self-referential, and the real-world benchmark partially contradicts the broad claim. The authors already acknowledge the BBOB limitation in §5.3, so the fix is within the manuscript's scope: add misspecification experiments or soften the claims, and specify the input normalization. The missing normalization description is easy to repair. This is a solid GECCO companion candidate after revision, and the paper is honest about several limitations in its conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it combines beta CDF input warping with affine domain transfer for surrogate models, and works out the optimization details carefully, including Riemannian gradients for rotations on SO(d) and a CMA-ES variant for non-differentiable surrogates. The gradient derivations in Eqs. (5)-(14) are correct, and the ablation study comparing beta-only, affine-only, and full transfer is a nice addition. The authors also ship code and supplementary material, and they are unusually candid in Section 5.3 about the limits of their benchmark design. That honesty earns credit.\n\nThe big soft spot is real and is exactly what the stress-test note flags: the BBOB targets are constructed inside the method's own model family, so the main synthetic results only show parameter recovery, not robustness to misspecification. The paper acknowledges this explicitly, but the abstract and introduction still claim the transferred surrogate \"significantly outperforms\" scratch training in data-scarce scenarios. That is too strong. On the out-of-family ABS benchmark, the full method beats affine-only on several instance pairs, but scratch training wins on instance3, and the paper's own Figure 23 shows mixed gains. So the evidence supports a narrower claim: the method helps when the transformation is close to beta-CDF-plus-rotation, and can hurt when it is not.\n\nA second issue is the undocumented input normalization. The beta CDF is defined on [0,1], yet the experiments sample from [-5,5]^d. The authors never say how inputs are mapped to the unit interval before applying phi. This is a small fix, but it is needed for reproducibility. The fact that the full method sometimes underperforms beta-only or affine-only in the ablation also deserves more discussion; it suggests the optimization landscape with extra parameters is harder, which is worth stating as a design trade-off rather than leaving it as a side observation.\n\nOverall, this is a solid incremental contribution, not a paradigm shift. The math is sound, the real-world benchmark is a genuine attempt at out-of-family validation, and the limitations are honestly disclosed. It deserves serious peer review, but the authors should be asked to clarify the input normalization, soften the abstract, and ideally add a misspecification experiment where the target is generated by a different warping family.","headline":"A useful incremental extension to surrogate transfer, with honest caveats but a circular synthetic benchmark and an overstated abstract.","tokens_in":37854,"tokens_out":1612,"would_cite":true,"duration_ms":17676,"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":"A pre-trained surrogate transfers to a new task by fitting a nonlinear input warp, a rotation, and a translation from only a handful of target points.","keywords":["transfer learning","Gaussian process regression","input warping","beta CDF","affine transformation","BBOB benchmark","low-data regression","surrogate model"],"falsifier":"Construct a target $f_T(x)=f_S(W\\phi(x)+v)$ with $\\phi$ deliberately outside the $\\beta$-CDF family, for example a non-monotonic coordinate map such as a sine perturbation, and compare the transferred and scratch-trained Gaussian processes on a 20-point transfer set; if the transferred model does not beat the scratch model on average, the claim that the $\\beta$-CDF parameterization captures the relevant nonlinearity is refuted.","tokens_in":36814,"feed_emoji":"🔁","tokens_out":5357,"duration_ms":48197,"temperature":0.7,"pith_summary":"This paper tries to establish that a surrogate model trained on one function can be reused on a related function by re-parameterizing its input as the composition of a coordinatewise beta cumulative distribution function (a smooth, monotone warp) with a rotation and a translation, then fitting the warp and affine parameters by minimizing mean squared error on a small set of target points. The motivation is that previous transfer methods assume only affine relations between source and target, which is too restrictive for real engineering problems. On the BBOB synthetic benchmark and an automotive braking benchmark, the transferred Gaussian process is claimed to be more accurate than both the original source surrogate and a Gaussian process retrained from scratch on the same transfer data, especially when only 20 to 40 target samples are available. The paper also shows that this advantage shrinks as transfer data grows and can reverse for very rugged or multimodal functions, so the method is a low-data tool rather than a large-data one.","feed_headline":"Fit a warped map, transfer a surrogate from 20 target points","feed_subtitle":"A beta-CDF warp plus rotation lets a pre-trained Gaussian process beat retraining when target data is scarce.","key_machinery":"The central object is the re-parameterized predictive map $\\hat{f}_S(W\\phi(x;\\theta)+v)$, where $\\phi$ applies a $\\beta$ CDF $\\int_0^{x_i} u^{\\alpha_i-1}(1-u)^{\\beta_i-1}/B(\\alpha_i,\\beta_i)\\,du$ to each coordinate, $W$ is a rotation matrix in $SO(d)$, and $v$ is a translation. Fitting $(\\theta,W,v)$ by minimizing MSE on the transfer set is carried out with Riemannian gradient descent for differentiable surrogates, which projects the Euclidean gradient of the rotation onto the tangent space of $SO(d)$ and takes steps along geodesics via the exponential map, and with CMA-ES over a flat $\\mathfrak{so}(d)$ representation for non-differentiable surrogates. The warp reshapes the source model's contours to match the target, and for fixed-kernel Gaussian processes it effectively materializes a non-stationary autocorrelation function.","core_discovery":"The central claim is that modeling the source-to-target relation as $f_T(x)=f_S(W\\phi(x)+v)$, with $\\phi$ a coordinatewise $\\beta$ CDF and $W\\in SO(d)$, turns a pre-trained surrogate $\\hat{f}_S$ into an accurate model of $f_T$ after fitting only the parameters $W$, $v$, and the $\\beta$ shape parameters on a tiny transfer set. The paper reports that with 20 transfer samples on 2D BBOB problems the transferred model beats a Gaussian process trained from scratch on most functions, and in 10D it remains strongly superior at 40 to 80 samples because scratch Gaussian processes collapse with so few points. On the automotive benchmark, the full warp-plus-affine model consistently outperforms an affine-only transfer baseline, and it beats scratch training for transfer sets smaller than roughly 30 points, with transfers involving one particular automotive instance identified as a failure case where the source-target relation is too intricate to capture.","pith_inferences":["A natural extension is to replace the beta CDF with a cheaper parametric warp such as the Kumaraswamy CDF, which the paper lists as future work and which would preserve the optimization setup while reducing cost.","The results suggest an active-learning regime: if each transferred sample is chosen to maximize disagreement between candidate warps, even fewer than 20 points might suffice, though the paper does not test this.","Boundary effects from rotation and translation can map out-of-domain regions into the target domain; the paper's in-domain ablation suggests that restricting to mapped-back transfer data sometimes helps, so a penalty discouraging out-of-domain mappings could improve robustness on real problems.","Because the method only re-parameterizes inputs, it is agnostic to the surrogate class; the same fitted warp should transfer random forests or neural networks provided parameter optimization is handled by a derivative-free method."],"forward_implications":["With only 20 target samples, transferred Gaussian processes beat scratch-trained Gaussian processes on most 2D and 5D BBOB functions and on nearly all 10D functions.","In 10D, scratch-trained Gaussian processes have very high SMAPE at 40 to 80 samples, so the transferred model's advantage is largest exactly where data is scarcest.","The advantage erodes as transfer data grows; by 80 to 200 samples, scratch training becomes competitive or better, making the method a low-data technique rather than a large-data one.","The full warp-plus-affine method outperforms affine-only transfer on the automotive benchmark, indicating that nonlinear warping captures source-target relations that affine maps miss.","Transfer fails or underperforms on rugged, multimodal functions such as F16 and F23-F24, where the original surrogate is already inaccurate, so the benefit depends on the source model having reached a baseline level of accuracy."],"supporting_citations":[{"why":"Defines the affine-only transfer baseline that this work extends by adding beta-CDF warping.","marker":"[23]"},{"why":"Introduces beta-CDF input warping, the nonlinear warp family used here.","marker":"[37]"},{"why":"Supplies the real-world automotive ABS-braking benchmark used to validate the method.","marker":"[39]"},{"why":"Shows affine-only transfer on synthetic and automotive benchmarks and motivates the need for nonlinear transfer.","marker":"[24]"},{"why":"Provides the CMA-ES optimization routine used to fit parameters for non-differentiable surrogates.","marker":"[17]"},{"why":"Supplies the automatic hyperparameter tuning routine used to set learning rates and schedules for the transfer optimization.","marker":"[22]"}],"fun_headline_variants":["Warp and rotate: transfer a surrogate from just 20 points","Beta-CDF warp beats scratch surrogate on tiny data","Surrogate transfer via beta warp: 20 points win","Domain warping plus affine: transfer with scarce data","Transfer pretrained surrogates with a warp and a twist"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the target is the source composed with a rotation, a translation, and a coordinatewise beta-CDF warp; if the true relation between source and target does not lie in this family, transfer can underperform training from scratch on the same data.","fun_headline_variants_meta":{"raw":{"variants":["Warp and rotate: transfer a surrogate from just 20 points","Beta-CDF warp beats scratch surrogate on tiny data","Surrogate transfer via beta warp: 20 points win","Domain warping plus affine: transfer with scarce data","Transfer pretrained surrogates with a warp and a twist"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000349,"raw_usage":{"total_tokens":1901,"prompt_tokens":935,"completion_tokens":966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":894}},"tokens_in":551,"tokens_out":966,"duration_ms":9344,"temperature":1.0,"reasoning_tokens":894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:50:59.122056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a target $f_T(x)=f_S(W\\phi(x)+v)$ with $\\phi$ deliberately outside the $\\beta$-CDF family, for example a non-monotonic coordinate map such as a sine perturbation, and compare the transferred and scratch-trained Gaussian processes on a 20-point transfer set; if the transferred model does not beat the scratch model on average, the claim that the $\\beta$-CDF parameterization captures the relevant nonlinearity is refuted.","supporting_citations":[{"cited_title":"Kononova","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world automotive ABS-braking benchmark used to validate the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces beta-CDF input warping, the nonlinear warp family used here."},{"cited_title":"Transfer Learning of Surrogate Models via Domain Affine Transformation Across Synthetic and Real-World Benchmarks","cited_arxiv_id":"2501.14012","evidence_quote":"Shows affine-only transfer on synthetic and automotive benchmarks and motivates the need for nonlinear transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the automatic hyperparameter tuning routine used to set learning rates and schedules for the transfer optimization."}],"review_version":1}