{"id":"8daa4949-a42d-4825-bfdf-eb242dec3549","arxiv_id":"2501.19161","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A locality-aware surrogate loss derived from the gradient theorem improves gradient estimation and query efficiency in black-box optimization.","lead":"GradPIE is a new training loss that makes surrogate models for black-box optimization match the true function's local gradients, reducing the number of expensive simulator queries needed. The method shows consistent gains on oscillator, circuit, and optical design tasks, and could speed up physics and engineering optimization loops.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equations (8)–(9) prove alignment for the ideal pair loss (5), but the optimized loss (10) is a k-NN empirical average; with K << D (OWMS: D=3600), it constrains the Jacobian on only K directions per point, so minimizing (10) does not imply Jacobian alignment.","rationale":"The paper's central claim is that minimizing GradPIE loss, as computed with k-NN in Eq. (10), corresponds to minimizing the Jacobian difference between surrogate and black-box function. The derivation in Eqs. (7)-(9) supports this for the idealized pair loss (5) under a small-distance assumption and a continuous density of pairs. The scalable loss (10) is an empirical average over K nearest neighbors, and no analysis connects it to the idealized loss. The reader identified the high-dimensional distance issue, which is valid. My stress test adds a sharper, more structural gap: even with small distances, the finite-K loss is a sum over at most K directions per point. In dimension 3600 with K on the order of 5-10, the loss cannot distinguish Jacobian perturbations that vanish on those K directions but are large elsewhere. Therefore the theoretical guarantee, as stated, does not apply to the algorithm that is actually run. This is load-bearing because the paper sells the theory as the reason the method works; without it, the empirical improvements, while plausible, lack the claimed mechanistic explanation. The proposed synthetic test would settle the issue by checking whether minimizing Eq. (10) actually yields the true Jacobian when the displacement span is low-rank. I do not change the reader's CONDITIONAL verdict because the empirical results may still hold and the paper could be revised to state a weaker claim or to add an analysis of the k-NN gap; the concern strengthens the conditions, it does not overturn the conditional recommendation.","tokens_in":14421,"tokens_out":10079,"duration_ms":106784,"concrete_test":"On a synthetic quadratic F(x) = 0.5 * x^T M x in D=100, with N=200 random training points and K=5, minimize Eq. (10) over a linear surrogate parameterized by A. Compare the minimizer A* to the true Jacobian M: if ||(A* - M) v|| is substantial for a random unit vector v while the loss is near zero, and the span of the displacement vectors has dimension at most K, then Eq. (10) does not enforce Jacobian alignment, and the theory of Eqs. (8)-(9) does not transfer to Eq. (10).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 derives the Jacobian-alignment guarantee for the pair loss L_GradPIE in Eq. (5): under ||x - x'|| <= epsilon, Eq. (9) shows L ≈ C * p(x,x) * ||J_F_hat - J_F||_row, so minimizing it aligns Jacobians. Section 4 replaces the full pair expectation with the k-NN empirical loss in Eq. (10). The paper never bounds the error between (10) and (5)/(9). Two gaps appear at once. (i) In the OWMS task, D=3600 but N=1000, so nearest-neighbor distances are not small; the linearization in Eq. (8) has an uncontrolled O(epsilon^2) error. (ii) Even when distances are small, the per-point loss in (10) with K neighbors penalizes the Jacobian difference A = J_F_hat - J_F only on K displacement vectors v_k = x'_k - x. Since the term is |A v_k|_1, any component of A orthogonal to span{v_1,...,v_K} is invisible to the loss. For K << D, a random direction v has expected squared projection K/D onto this span; for D=3600 and K=10 that is about 0.003. Thus the empirical loss can be arbitrarily small while ||A||_row is large in most directions. The claimed equivalence in the abstract and Section 3 is therefore not inherited by the scalable loss actually optimized. The reported gains may come from the local sampling/retraining schedule or the network's inductive bias rather than from the stated Jacobian-alignment mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a surrogate-model training loss, GradPIE, for gradient-based black-box optimization. Starting from the gradient theorem, the authors replace the black-box function inside a path integral with the surrogate and obtain a pair-loss objective (Eq. 5), then make it scalable by using k-nearest-neighbor pairs (Eq. 10). They claim that minimizing this loss aligns the Jacobian of the surrogate with the Jacobian of the black-box function, and they evaluate the method in offline and online active black-box optimization on three tasks: coupled nonlinear oscillators (CNON), an analog OpAmp circuit, and a 3600-dimensional optical wave manipulation system (OWMS). The experiments report consistent improvements over a base surrogate trained with MAE and over Bayesian optimization and random search under limited query budgets.","tokens_in":14839,"tokens_out":6816,"duration_ms":73511,"significance":"If the Jacobian-alignment claim were established, GradPIE would be a simple and cheap modification to surrogate training for black-box optimization, and the paper would make a useful practical contribution. The work has real strengths: the loss is derived from an integral identity rather than from fitted constants; the CNON experiment directly compares estimated gradients with exact gradients; and the empirical evaluation spans three heterogeneous tasks, including a high-dimensional optical system. However, the theoretical relation is established only for an ideal pair loss, not for the k-NN loss that is actually optimized, and the high-dimensional experiments are conducted in a regime where the small-displacement assumption is questionable. The empirical gains are promising but, without an ablation or a formal bound, the paper's central mechanism remains conditional.","major_comments":[{"comment":"The Jacobian-alignment conclusion of Eq. (9) applies to the ideal pair loss in Eq. (5), but the optimized loss in Eq. (10) is a k-NN empirical average, and the paper provides no bound linking the two objectives. For a fixed x, the per-point term in Eq. (10) penalizes A = J[Fhat] - J[F] only on the K displacement vectors v_k = x'_k - x; any component of A orthogonal to span{v_1, ..., v_K} is invisible to the loss. In the OWMS task (D = 3600, N = 1000), K is necessarily much smaller than D, so the loss can be arbitrarily small while the row-norm Jacobian discrepancy is large in most directions. The claimed equivalence in the abstract and Section 3 is therefore not inherited by the scalable loss actually used in training.","section":"Sections 3-4, Eqs. (5), (9), (10)"},{"comment":"The derivation of Eq. (9) assumes paired inputs are close, with ||x - x'|| <= epsilon, and replaces the line integral of A(u) by A(x)(x' - x). The resulting O(epsilon^2) Taylor error is uncontrolled. In the OWMS experiment, 1000 training points in a 3600-dimensional space are sparse, so typical nearest-neighbor distances are not small, and the paper gives no evidence that the k-NN pairs used in Eq. (10) satisfy the epsilon condition. Without reporting pairwise distances or bounding second derivatives of F and Fhat, the theoretical justification does not apply to the high-dimensional regime on which the paper's main empirical claim rests.","section":"Section 3, Eq. (8)"},{"comment":"The reported optimization gains are not uniquely attributable to GradPIE. The online procedure in Algorithm 1 simultaneously introduces local sampling (Ns), rank selection, and periodic retraining; for OpAmp and OWMS, no Ns = 0 ablation is reported, so the improvement relative to the base model could be driven by the local-sampling schedule or by the network's inductive bias rather than by Jacobian alignment. Since the theoretical mechanism is in question, an ablation that isolates the loss function under identical sampling schedules, or per-iteration gradient-alignment error on all tasks, is needed to support the paper's central claim.","section":"Section 5, Algorithm 1, Table 1"}],"minor_comments":[{"comment":"The nearest-neighbor count K is a central hyperparameter, but the paper does not report the K values used for the OpAmp and OWMS tasks, nor the standard deviation sigma and Ns used in local sampling; please add a hyperparameter table for reproducibility.","section":"Section 5, Table 3"},{"comment":"The stopping condition 'if LGradPIE < epsilon then Break' is scale-dependent because the loss in Eq. (10) is an unnormalized sum of L1 differences; the threshold should be normalized or removed.","section":"Algorithm 1"},{"comment":"Figures 2f-2h report optimization curves without confidence intervals, and the text says the 'plots clearly show' superiority; given the stochasticity of the procedure, error bars or multiple-seed statistics are needed to support this statement.","section":"Figures 2f-2h"},{"comment":"The statement that an MSE formulation 'would yield similar results' is unsupported; either provide an experiment or clearly label this as a conjecture.","section":"Section 3"},{"comment":"There are minor typographical issues, including the broken character in 'naively' and some informal usage of 'O(.)' in Table 2; these should be cleaned up.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The empirical component is solid enough to be worth publishing after revision, but the theoretical claim is currently overclaimed relative to what is proved. I would ask the authors to either provide a quantitative bound for the k-NN loss or explicitly soften the Jacobian-alignment claim to a heuristic motivation. I see no novelty-disclosure issue; the concern is purely about matching the theoretical statement to the algorithm and experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe GradPIE loss is a simple, useful idea: train a surrogate to reproduce local differences F(x')-F(x) instead of absolute values. That is a genuinely sensible way to get better gradient estimates without extra black-box queries, and the paper shows consistent improvements over a MAE-trained base model across three real tasks (CNON, OpAmp, OWMS). The CNON experiment with exact gradients is a nice sanity check that the loss actually helps gradient estimation, not just optimization curves.\n\nThe soft spot is real and load-bearing for the theory. Section 3 proves that the ideal pair loss in (5) approximates the row-norm difference of Jacobians when pairs are close. But the optimized loss in (10) is a k-NN empirical average. For the OWMS task, D=3600 with N=1000, the nearest-neighbor distances are not small and the linearization in (8) is uncontrolled. More importantly, even with small distances, (10) with K=10 (say) constrains the Jacobian difference on only 10 directions per point; the other ~3590 directions are invisible. The paper never bounds the gap between (10) and (5), so the claim that minimizing (10) aligns Jacobians is not established. This is not a minor footnote; it is the paper's central theoretical contribution. The reported optimization gains may come from the local sampling/retraining schedule or from the MLP's inductive bias, not from the stated mechanism.\n\nThat said, the paper is not a waste of time. The empirical finding that a difference-matching loss improves surrogate gradients is worth reporting, and the method is trivially easy to implement. The weaknesses are addressable: release code and data, add the gap analysis or tone down the theory, compare against L-GSO (the closest prior local-generative-surrogate method), and report the hyperparameters (K, sigma, Ns). The 3600-D OWMS result without measuring neighbor distances is the thing I would most want fixed.\n\nThe reader's conditional verdict is fair. The stress-test note about direction-blindness in high dimensions is correct and should be part of the review discussion. I would send this to peer review: the idea deserves a serious referee even if the current framing oversells the theory. With a revised theory section and honestly stated limitations, it could become a solid, citable piece.\n\nBest,\n[Your name]","headline":"A clean difference-matching surrogate loss with consistent empirical gains, but the theory overreaches: the Jacobian-alignment guarantee is proven only for the full pair loss, not the k-NN loss actually optimized.","tokens_in":15328,"tokens_out":3359,"would_cite":true,"duration_ms":32548,"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":"Training a surrogate on nearest-neighbor value differences aligns its Jacobian with the black-box function, giving better gradients for optimization.","keywords":["black-box optimization","surrogate model","gradient estimation","GradPIE loss","locality-aware training","active learning","gradient theorem","nearest-neighbor pairs"],"falsifier":"On a smooth test function with exactly computable Jacobians, train two surrogates with the GradPIE loss using small-radius and large-radius neighbor pairs; if the large-radius model shows decreasing loss while the Jacobian error $\\|J[\\hat{F}] - J[F]\\|$ stays flat or grows, the stated equivalence fails outside the small-radius regime. A cheaper check is to measure actual nearest-neighbor distances in the 3600-dimensional OWMS dataset: if typical distances exceed the linearization scale, the theoretical guarantee does not apply.","tokens_in":14261,"feed_emoji":"🎯","tokens_out":6516,"duration_ms":58026,"temperature":0.7,"pith_summary":"The paper tries to establish that a simple change to surrogate training produces substantially better gradient estimates for black-box optimization. Instead of matching predicted outputs pointwise, the proposed GradPIE loss compares, for each input and its nearest neighbors, the difference in predicted outputs with the difference in observed black-box outputs. The paper argues, via the gradient theorem, that minimizing this pairwise loss is equivalent to minimizing the gap between the Jacobians of the surrogate and the black-box function. Since optimization uses surrogate gradients in the backward pass, better Jacobian alignment translates into more efficient search over non-differentiable simulators. Experiments on coupled nonlinear oscillators, an operational amplifier circuit, and a 3600-dimensional optical wave-shaping system support this: the locality-aware surrogate outperforms the same architecture trained with MAE or MSE, and it reaches the baseline's quality with roughly 40-70% fewer black-box queries.","feed_headline":"Pairwise loss aligns surrogate gradients with black-box functions","feed_subtitle":"Minimizing the GradPIE loss over nearby samples cuts black-box queries by up to ~70 percent in tested tasks.","key_machinery":"The load-bearing object is the GradPIE loss itself: $\\mathcal{L}_{\\mathrm{GradPIE}}(\\theta) = \\mathbb{E}_{x \\sim \\mathcal{D}} \\frac{1}{K} \\sum_{k=1}^{K} \\left| (F(x) - F(x'_k)) - (\\hat{F}(x;\\theta) - \\hat{F}(x'_k;\\theta)) \\right|$, where $x'_k$ is the $k$-th nearest neighbor of $x$. Unlike a pointwise MAE or MSE loss, this loss compares value differences, so an additive constant in either function cancels; computed locally, it forces the surrogate to reproduce the same relative changes in output that the black-box shows. The gradient theorem then converts this pairwise consistency into an integral of Jacobian differences, which is the mechanism behind the claim.","core_discovery":"The central claim, in the paper's own terms, is that minimizing the GradPIE loss in Eq. (5), computed over k-nearest-neighbor pairs in Eq. (10), corresponds to minimizing the difference between the Jacobians of the surrogate and the black-box function. Starting from the gradient theorem, the loss is rewritten as an integral of the Jacobian difference along the curve between paired inputs; under a locality assumption $\\|x - x'\\| \\le \\epsilon$, this integral is proportional to the row-wise norm of $J[\\hat{F}](x;\\theta) - J[F](x)$. Hence GradPIE-trained surrogates provide gradient estimates that better match the true gradients, and the paper demonstrates this both by direct gradient-error measurements and by optimization results on three real-world tasks.","pith_inferences":["Going beyond the paper: GradPIE could be used as an auxiliary loss or regularizer in any surrogate-training pipeline, since it only needs value differences; the paper limits its experiments to MLP surrogates and three tasks.","Going beyond the paper: weighting neighbor pairs by inverse distance, or annealing the neighbor radius, is a natural extension that might restore the Jacobian-alignment guarantee when data is scarce; the paper does not test this.","Going beyond the paper: because GradPIE ignores global offsets, it is closely related to contrastive or rank-based learning of scalar functions; that link might transfer the method to problems where only relative comparisons, not absolute values, are observed."],"forward_implications":["Training a surrogate with GradPIE instead of MAE or MSE improves gradient estimation on the same dataset, with relative error reductions over 20% and cosine-similarity gains over 8% in the oscillator experiments.","In active black-box optimization, the locality-aware model reaches the base model's 200-iteration performance with roughly 40-70% fewer black-box calls across the three tasks.","The loss works in both offline and online settings, so it can be applied when the simulator can be queried during optimization and when only a fixed dataset exists.","Simple MLP surrogates trained this way remain effective in a 3600-dimensional search space, indicating the method does not rely on specialized architectures."],"supporting_citations":[{"why":"Supplies the gradient theorem used to derive GradPIE and the Jacobian-alignment equivalence.","marker":"Williamson & Trotter, 2004"},{"why":"Establishes the local generative surrogate approach to black-box optimization that the paper builds on and compares against.","marker":"Shirobokov et al., 2020"},{"why":"Provides the coupled nonlinear oscillator network task and the context of training physical neural networks with surrogate gradients.","marker":"Wright et al., 2022"},{"why":"Provides the differentiable optics simulator used for the optical wave manipulation system experiment.","marker":"Filipovich & Lvovsky, 2024"},{"why":"Defines the OpAmp figure-of-merit used as the objective in the analog circuit benchmark.","marker":"Dong et al., 2023"}],"fun_headline_variants":["GradPIE loss cuts black-box queries by up to 70%","Locality-aware surrogates match black-box gradients","Pairwise loss aligns gradients to reduce queries","GradPIE: train surrogates that see true gradients","Gradient-path loss reduces black-box optimization queries"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes the black-box function is differentiable enough for the gradient theorem to apply and that paired inputs are truly close, with $\\|x - x'\\| \\le \\epsilon$; in high-dimensional problems with limited data, nearest neighbors can be far apart, and then minimizing GradPIE is not guaranteed to align the Jacobians.","fun_headline_variants_meta":{"raw":{"variants":["GradPIE loss cuts black-box queries by up to 70%","Locality-aware surrogates match black-box gradients","Pairwise loss aligns gradients to reduce queries","GradPIE: train surrogates that see true gradients","Gradient-path loss reduces black-box optimization queries"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3283,"prompt_tokens":864,"completion_tokens":2419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2338}},"tokens_in":480,"tokens_out":2419,"duration_ms":17143,"temperature":1.0,"reasoning_tokens":2338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:06:01.911574+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a smooth test function with exactly computable Jacobians, train two surrogates with the GradPIE loss using small-radius and large-radius neighbor pairs; if the large-radius model shows decreasing loss while the Jacobian error $\\|J[\\hat{F}] - J[F]\\|$ stays flat or grows, the stated equivalence fails outside the small-radius regime. A cheaper check is to measure actual nearest-neighbor distances in the 3600-dimensional OWMS dataset: if typical distances exceed the linearization scale, the theoretical guarantee does not apply.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gradient theorem used to derive GradPIE and the Jacobian-alignment equivalence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the local generative surrogate approach to black-box optimization that the paper builds on and compares against."},{"cited_title":"G., Onodera, T., Stein, M","cited_arxiv_id":null,"evidence_quote":"Provides the coupled nonlinear oscillator network task and the context of training physical neural networks with surrogate gradients."}],"review_version":1}