{"id":"5094ef92-64ec-4111-97c3-dbfbffc61380","arxiv_id":"2608.06894","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ESO conditions spectral mode mixing on local pairwise variation statistics, improving neural operator accuracy on PDEs with sharp local structures.","lead":"This paper introduces a neural network for solving partial differential equations that adjusts its calculations based on local differences between neighboring grid points, not just each point's own value. It reports lower error than existing methods on nine standard PDE benchmarks, especially near sharp material boundaries and high-gradient flow structures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on a single-run comparison; the smallest reported margin (Navier-Stokes) is within likely seed noise, so the headline needs variance reporting or it is not supported.","rationale":"I agree with the reader's weakest assumption. The architecture is plausible, the ablations are internally consistent, and there is no obvious mathematical contradiction. But the paper's headline is an empirical dominance claim, and the reported margins on two benchmarks are smaller than typical seed-level variation for operator learning. The requested check is concrete and decisive: it directly tests whether the central claim survives re-running with variance. If the margins persist, I would accept the SOTA claim; if not, the paper should be conditional on adding error bars and tuned baselines. Since this is exactly the reader's conditional, the verdict does not change.","tokens_in":15235,"tokens_out":4410,"duration_ms":53343,"concrete_test":"Using the released repository, rerun ESO and LRSA (the strongest baseline) on Navier-Stokes and Darcy Flow with at least five random seeds each, identical hyperparameters and a verified equal parameter budget; report mean ± std of relative L2. If the Navier-Stokes gap (4.77e-2 vs 4.84e-2) is within one pooled standard error, or if ESO does not win on all seeds, the SOTA claim is not established. Also tabulate parameter counts for every model in Tables 2/3 to confirm the 'comparable parameter budgets' assertion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: ESO 'consistently achieves state-of-the-art performance' on nine PDE benchmarks. The theory (Theorems 1-2) is conditional and does not by itself establish superiority, so the tables carry the argument. The paper reports one run per model and asserts that all baselines are matched with ESO under comparable parameter budgets and consistent training configurations, but it gives no seed variance, no standard deviations, and no parameter counts in Tables 2/3. The decisive gap is small: on Navier-Stokes, ESO is 4.77e-2 vs LRSA 4.84e-2 (a difference of 7e-4 in relative L2, about 1.4%); on Heat Transfer, 9.75e-5 vs 1.49e-4. These margins are of the size that commonly appears as run-to-run scatter in neural operator training. If baseline tuning was not fully equalized (e.g., LRSA was taken from a prior paper rather than retuned to the same budget), or if seed variance exceeds the margin, the 'consistent SOTA' conclusion collapses even though the architecture may be sound. This is a measurement-procedure concern, not an attack on the method's validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Edge-Conditioned Spectral Operator (ESO), a neural operator that modulates spectral mode selection using local pairwise variation statistics (PVMM) and a task-adaptive Physics-Aware Reweighting (PAR) loss. ESO is evaluated on nine PDE benchmarks spanning structured and unstructured meshes, where it reports the lowest relative L2 error on every task, supported by ablations, qualitative visualizations, and a resolution-generalization experiment. The theoretical section claims a no-worse approximation guarantee and a separation result for point-conditioned versus edge-conditioned gates.","tokens_in":15524,"tokens_out":5111,"duration_ms":57273,"significance":"The central idea is natural and potentially useful: conditioning spectral modal gates on local pairwise variation statistics is a modest but plausible extension of point-conditioned spectral operators, and the evaluation scope is broad, covering both structured and unstructured mesh problems. The code release is a concrete asset, and the ablations clearly separate the contributions of PVMM and PAR. However, the headline claim of consistent state-of-the-art performance is not yet fully established: the empirical comparison is single-run, with no variance reporting, and the closest margins are small; the theoretical results are class-inclusion and existence statements rather than predictive guarantees. If the empirical comparison is strengthened with seed-level statistics and verifiable parameter-budget matching, the paper would make a useful contribution to operator learning.","major_comments":[{"comment":"The central claim of consistent state-of-the-art performance rests on single-run relative L2 numbers. On Navier–Stokes (Table 2), ESO is 4.77e-2 versus LRSA 4.84e-2, a margin of about 1.4%, which is small relative to typical seed-to-seed variation in neural operator training. Without multiple seeds, standard deviations, and a significance test, the headline claim is not supported to the usual empirical standard. Please report seed-level results, mean ± std, and ideally paired significance tests for the closest comparisons.","section":"Experiments, Tables 2 and 3"},{"comment":"The statement that \"all baselines are matched with ESO under comparable parameter budgets and consistent training configurations\" is not verifiable as written: Tables 2 and 3 contain no parameter counts, and the paper does not state whether each baseline was retrained under the same schedule or whether reported numbers are taken from original publications. Because the closest baseline LRSA is only marginally worse on several tasks, the fairness of the comparison is load-bearing. Provide a parameter-count table, the training schedule used for each baseline, and a statement of which numbers are reproduced versus cited.","section":"Experimental Setups, Metrics and Appendix Tables 7/8"},{"comment":"The theoretical analysis is conditional rather than predictive. Theorem 1 (Eq. 21) follows from the class inclusion F_base ⊆ F_edge by setting the edge coefficient to zero (Appendix Eq. 25 with αE = 0), so it only shows no-worse approximation, not superiority. Theorem 2 assumes h_i(a) = h_i(ã) in Eq. (22), but in a multi-layer spectral operator the latent h_i has already passed through global mixing, so this premise is not guaranteed for the actual architecture; the constructed interpolation r(c) in the appendix is defined on descriptor pairs, not on the original input space, and the proof does not show that the target π* belongs to the class of gates realizable by ESO. Please state explicitly that the theorems are existence/class-inclusion results and avoid presenting them as explaining the empirical gains.","section":"Theoretical Perspective and Appendix Proofs, Theorems 1–2"}],"minor_comments":[{"comment":"The MG-TFNO citation is incomplete: \"Kossaifi, J.; ... ????\" appears without a year or venue; supply the full bibliographic information.","section":"References"},{"comment":"The appendix renumbers the main-text Theorems 1 and 2 as Theorems 3 and 4; use a single numbering scheme throughout the paper.","section":"Appendix Proofs"},{"comment":"The expression for \\tilde w_i would be clearer with explicit parentheses around (\\bar{s}^τ + ε); as printed, the precedence is ambiguous.","section":"Task-adaptive Physics-Aware Reweighting, Eq. (14)"},{"comment":"Several entries in the \"Emphasized region\" column are sentence fragments (e.g., \"vortices and shear/high-gradient regions\"); rewrite them as complete noun phrases or short clauses.","section":"Appendix PAR Implementation Details, Table 6"},{"comment":"The tensor contraction in R((Φ^E)ᵀVᵗ) should be spelled out, since R is a three-index tensor and the matrix notation is ambiguous as written.","section":"Method, Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the headline SOTA claim may not survive multi-seed evaluation. I would ask the authors to add seed-level statistics and parameter counts; if the close margins persist with proper variance reporting, the paper should be acceptable. The theory section should also be toned down from \"separation\" language to conditional/existence language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental neural operator paper whose headline claim—SOTA on nine PDE benchmarks—is probably true but not yet supported as stated, because the entire comparison is single-run and the decisive margins on some tasks are smaller than typical seed noise. The architectural idea is simple and sensible: instead of conditioning spectral mixing only on the point's own latent feature, also feed it local mean-difference and squared-difference statistics from its neighbors. That is a genuine, if modest, extension of HPM-style adaptive spectral gates. The PAR loss reweighting is a standard trick, but it is disclosed and it helps a bit in the ablations. The paper also ships code and gives detailed training configurations, which is more than many papers in this area do.\n\nWhat is actually new: the PVMM gate is a real departure from point-centered gating. The ablations show that adding D and Q each helps, and that the full combo beats the point-only gate on Darcy and Navier–Stokes. The resolution-generalization plot is a nice addition. The theory, however, is not a strength. Theorem 1 is just class inclusion (set the edge coefficient to zero), so the 'no-worse' conclusion is definitional. Theorem 2 is an existence argument with a contrived condition; it does not prove that a specific ESO achieves the separation, only that a sufficiently expressive gate could. That is fine as intuition, but it should be labeled as such.\n\nThe soft spot is the empirical comparison. The paper states that baselines are matched with ESO under comparable parameter budgets and consistent training configurations, but no parameter counts or standard deviations appear in Tables 2 and 3. On Navier–Stokes the margin over LRSA is 4.77e-2 vs 4.84e-2—about 1.4% relative—which is well within the run-to-run scatter I usually see for these models. On Heat Transfer the margin is larger in relative terms, but still a single run. So the 'consistently achieves SOTA' claim needs either multiple seeds or a tighter protocol. I don't think this is fatal: the ablations show the architecture helps, and the gain on Darcy is substantial. But the headline currently overstates the evidence.\n\nWho this is for: anyone working on neural operators, especially spectral methods, will find the edge-conditioning idea worth a look. It is not a breakthrough, but it is a useful building block. The paper deserves a serious referee; a competent reviewer should ask for variance reporting and a clearer statement of what was retuned versus taken from prior papers, and the theory section should be trimmed or reframed as remarks.\n\nMy recommendation: send it to peer review, but treat the SOTA claim as provisional. If the authors add seeds and the margins survive, it is a fine paper.","headline":"Solid incremental neural operator paper; the edge-conditioned gate is a real idea, but the SOTA claim outruns the single-run evidence.","tokens_in":16015,"tokens_out":2166,"would_cite":true,"duration_ms":23462,"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":"This paper claims that spectral PDE operators should gate their global mode mixing on local neighbor differences, and reports state-of-the-art accuracy on nine benchmarks with such an edge-conditioned operator.","keywords":["neural operators","PDE surrogate modeling","spectral methods","adaptive modal mixing","edge-conditioned gating","physics-aware reweighting","Darcy flow","unstructured meshes"],"falsifier":"Re-run the nine benchmark comparisons with multiple random seeds and report the mean plus standard deviation. If seed variation makes the reported margins overlap, such as the Navier-Stokes gap of 4.77e-2 for ESO versus 4.84e-2 for LRSA or the Pipe Turbulence gap of 4.32e-3 versus 4.89e-3, then the consistent-state-of-the-art claim would not be supported by the evidence as reported.","tokens_in":15041,"feed_emoji":"🌊","tokens_out":8075,"duration_ms":82602,"temperature":0.7,"pith_summary":"Spectral neural operators mix information globally in a truncated basis, but their adaptive variants decide how to weight each mode from the representation at a single point. This paper argues that such point-centered gating is blind to the local structures that actually control many PDE solutions, such as permeability jumps in Darcy flow and vortices in Navier-Stokes. To close that gap, the authors propose the Edge-Conditioned Spectral Operator (ESO), whose modal gate is conditioned on two local pairwise statistics: the signed average difference to neighboring points and the average squared difference. A second component, Physics-Aware Reweighting (PAR), sharpens the training loss in regions where a task-specific physical field changes abruptly. Across nine structured and unstructured PDE benchmarks, ESO reports the lowest relative $L^2$ error on every task, with the largest visual gains exactly in the sensitivity-highlighted regions.","feed_headline":"Edge-conditioned spectral gating wins on nine PDE benchmarks","feed_subtitle":"A pair of neighbor-difference statistics tunes spectral mode weights, cutting errors at permeability jumps and vortices.","key_machinery":"The load-bearing module is the Pairwise-Variation Modal Mixer (PVMM). For each point $x_i$, PVMM computes a signed shift statistic $D_{x_i}=\\frac{1}{|\\mathcal{N}(x_i)|}\\sum_{x_j\\in\\mathcal{N}(x_i)}(v_t(x_j)-v_t(x_i))$ and an unsigned contrast statistic $Q_{x_i}=\\frac{1}{|\\mathcal{N}(x_i)|}\\sum_{x_j\\in\\mathcal{N}(x_i)}(v_t(x_j)-v_t(x_i))^{\\odot 2}$, then a gating network maps the concatenation of the center feature, $D$, and $Q$ to $K$ spectral-mode logits. These logits are blended with the point-conditioned logits, softmaxed to form an edge-conditioned score $\\Pi^E$, and elementwise multiplied with the spectral basis $\\Phi$; the mixed output is $\\sigma(V_t W + \\tilde{\\Phi}^E R(\\tilde{\\Phi}^E)^\\top V_t)$. The neighborhoods are logical grid stencils on structured meshes and k-nearest-neighbor graphs in Laplace-Beltrami spectral coordinates on unstructured meshes. Separately, PAR turns a task-available physical field into a local sensitivity score, normalizes it, and uses it to reweight the relative-$L^2$ training loss, complementing the architectural change with explicit supervision in physically important regions.","core_discovery":"The central claim is that the dominant failure mode of adaptive spectral operators is not spectral resolution but the point-centered gate: when two inputs give the same latent feature at a node but differ in the neighboring field, a point-conditioned model is forced to give the same node-wise response, so it can be wrong by at least half the target gap on one input. ESO removes this indistinguishability by computing edge descriptors $D$ and $Q$ and letting them modulate the spectral basis; the paper's Theorem 1 shows the edge-conditioned class contains the point-conditioned class, so adding edges cannot hurt the best achievable risk, and its Theorem 4 states the separation lower bound for point-only gates. Empirically, the paper's claim is that this design consistently achieves state-of-the-art performance on nine PDE benchmarks, and that the gains concentrate near coefficient jumps, vortical structures, and deformation-localization regions.","pith_inferences":["Editorial extension: the separation theorem is stated for modal gates, but the same reasoning applies to any per-node prediction rule that ignores neighbor context, so graph convolutions and local attention with explicit edge features could inherit a similar argument, a direction the paper does not explore.","Editorial extension: a stress test not run in the paper would vary Darcy permeability contrast beyond the training range; if edge statistics carry the improvement, ESO's margin over the best point-centered baselines should widen as jumps sharpen and shrink as the field becomes smooth.","Editorial extension: PAR's sensitivity maps are built from task-specific physical fields assumed available during training, so in applications with noisy or misspecified fields the reweighting could emphasize artifacts; the reported gains may depend on clean benchmark quantities.","Editorial extension: the paper fixes the spectral basis and changes only the gate, so the benefit of edge conditioning under other global bases, such as wavelets or pure Laplace eigenfunctions, remains untested and is a natural next experiment."],"forward_implications":["ESO improves relative $L^2$ error on all nine benchmarks in Tables 2 and 3, with gains such as from 7.34e-2 to 4.77e-2 on Navier-Stokes and from 1.84e-4 to 9.75e-5 on Heat Transfer.","Because setting the edge coefficient to zero recovers the base point-conditioned spectral operator, the edge-conditioned class is at least as expressive, so adding edge conditioning cannot make the best achievable risk worse.","For inputs distinguished only by neighboring structure, any point-only modal gate carries a guaranteed error lower bound of $\\Delta/2$, whereas ESO can assign different mode weights to those inputs and is not subject to the bound.","Ablations on Darcy Flow and Navier-Stokes show that both $D$ and $Q$ contribute, and that PAR adds a further improvement; dropping either component raises error.","Resolution-generalization experiments on Darcy Flow show ESO's error growing more slowly than baselines beyond $85^2$ resolution, indicating that edge conditioning helps preserve fine-scale physics-sensitive structure."],"supporting_citations":[{"why":"Defines the Fourier spectral mixing operator that ESO extends and supplies the FNO backbone used throughout the comparisons.","marker":"Li et al. 2020"},{"why":"Supplies the point-conditioned adaptive spectral operator HPM that PVMM extends, the Laplace-Beltrami spectral basis convention, and the benchmark protocol.","marker":"Yue, Yang, and Zhu 2025"},{"why":"Introduces Laplace-Beltrami eigenfunction representation on manifolds and the NORM baseline, providing the spectral-coordinate space used for unstructured neighborhoods.","marker":"Chen et al. 2024"},{"why":"LRSA is the strongest baseline, second-best on all four structured tasks, so ESO's claimed state-of-the-art margin is measured against it.","marker":"Yang et al. 2026"},{"why":"Transolver is the best attention baseline on several structured tasks, especially Plasticity, and anchors the attention-based comparison.","marker":"Wu et al. 2024"},{"why":"GINO is the geometry-informed graph baseline that represents the graph-based family in the structured comparisons.","marker":"Li et al. 2023b"},{"why":"POD-DeepOnet is one of the strongest unstructured-mesh baselines and anchors the unstructured comparison.","marker":"Lu et al. 2022"},{"why":"Supplies the k-nearest-neighbor construction used to define unstructured neighborhoods in the mesh spectral-coordinate space.","marker":"Belkin and Niyogi 2003"}],"fun_headline_variants":["Edge-conditioned spectral operators sharpen PDE solutions near jumps","Using edges to tune spectral mixing beats point-only PDE models","Edge-aware spectral kernels win on nine PDE benchmarks","Neighbor differences guide spectral attention to sharpen PDE results","Edge-conditioning spectral operators: nine wins, sharper at jumps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that ESO is consistently state of the art assumes the baselines were given genuinely comparable parameter budgets and training configurations, and the paper reports a single run per model with no error bars, so the headline rests on the fairness and stability of the comparison procedure rather than on the architecture alone.","fun_headline_variants_meta":{"raw":{"variants":["Edge-conditioned spectral operators sharpen PDE solutions near jumps","Using edges to tune spectral mixing beats point-only PDE models","Edge-aware spectral kernels win on nine PDE benchmarks","Neighbor differences guide spectral attention to sharpen PDE results","Edge-conditioning spectral operators: nine wins, sharper at jumps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000537,"raw_usage":{"total_tokens":2591,"prompt_tokens":968,"completion_tokens":1623,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1545}},"tokens_in":584,"tokens_out":1623,"duration_ms":10857,"temperature":1.0,"reasoning_tokens":1545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:49:56.122616+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the nine benchmark comparisons with multiple random seeds and report the mean plus standard deviation. If seed variation makes the reported margins overlap, such as the Navier-Stokes gap of 4.77e-2 for ESO versus 4.84e-2 for LRSA or the Pipe Turbulence gap of 4.32e-3 versus 4.89e-3, then the consistent-state-of-the-art claim would not be supported by the evidence as reported.","supporting_citations":[],"review_version":1}