{"id":"08c26890-4479-4a32-b192-461049f23aa3","arxiv_id":"2605.24658","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"WLNO augments LNO with a parallel Haar wavelet branch and learnable gate to capture multi-scale spatial features, outperforming LNO on five PDE benchmarks especially those with sharp structures.","lead":"The paper introduces WLNO, a neural operator fusing Haar wavelet multi-scale decomposition with the Laplace Neural Operator via a gated fusion weight. A smart generalist might read it to understand how combining Laplace-domain dynamics with explicit spatial frequency handling can improve ML models for physical simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Performance gains lack ablation or variance controls to attribute improvement to wavelet branch","rationale":"The reader's weakest_assumption directly identifies the same attribution gap. Full-text access does not alter this because the abstract's claim of identical protocols is not supplemented by the required controls; the concern therefore remains load-bearing for the strongest_claim.","tokens_in":1819,"tokens_out":299,"duration_ms":15936,"concrete_test":"Retrain both WLNO and LNO on Burgers and 2D NS using the exact reported protocol but with 5 independent seeds; report mean relative L2 error plus std. Separately, ablate by fixing alpha_wav=0 throughout training and compare to full WLNO; if the ablated version matches LNO within error bars, the attribution holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the parallel Haar DWT branch (with learnable 1x1 convs per subband and sigmoid-gated alpha_wav) produces consistent outperformance over LNO on the five PDEs, especially multi-scale cases. The manuscript states identical hyperparameters, data, and protocols are used, yet supplies no ablation (e.g., alpha_wav fixed at 0 or wavelet branch removed) and no multi-seed statistics. Without these controls, observed deltas could arise from optimizer stochasticity, initialization, or minor unstated implementation differences rather than the wavelet-Laplace fusion.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces the Wavelet-Laplace Neural Operator (WLNO), which augments the Laplace Neural Operator (LNO) with a parallel single-level Haar DWT branch. The branch decomposes lifted features into LL/LH/HL/HH subbands, applies independent learned 1x1 convolutions per subband, reconstructs via inverse DWT, and fuses the result with the LNO core through a learnable sigmoid-gated scalar alpha_wav (initialized for small initial wavelet contribution). WLNO is evaluated on five PDE benchmarks (diffusion, Burgers, reaction-diffusion, Darcy flow, 2D Navier-Stokes) using identical hyperparameters, data, and protocols to LNO, with the claim of consistent outperformance that is most pronounced on problems with strong spatial multi-scale structure.","tokens_in":1956,"tokens_out":576,"duration_ms":28436,"significance":"If the reported gains are robustly attributable to the wavelet branch, the work would show that explicit multi-scale spatial decomposition can usefully complement Laplace-domain pole-residue modeling for neural operators. The adaptive gating mechanism and per-subband convolutions constitute a clean architectural extension that could be adopted in other frequency-domain operator frameworks.","major_comments":[{"comment":"Experiments section (and abstract claim of outperformance): the manuscript states that identical hyperparameters, training data, and evaluation protocols are used, yet supplies neither ablation results (e.g., alpha_wav fixed at 0 or wavelet branch removed) nor multi-seed statistics, error bars, or variance estimates. Without these controls the observed deltas cannot be confidently attributed to the wavelet-Laplace fusion rather than optimizer stochasticity, initialization, or minor implementation differences.","section":"Experiments section"},{"comment":"Method section (description of fusion): the claim that the model 'adaptively balances' the two branches throughout training rests on the learned alpha_wav, but no analysis is provided of its converged values, sensitivity to initialization, or correlation with performance gains on the multi-scale problems (Burgers, Navier-Stokes). This leaves the mechanistic explanation for the reported improvements under-specified.","section":"Method section"}],"minor_comments":[{"comment":"Abstract: asserts 'consistent outperformance' and 'most pronounced improvement' without any numerical values, table references, or error metrics, which reduces the abstract's standalone informativeness.","section":"Abstract"},{"comment":"Notation: the per-subband 1x1 convolution weights are described as 'independent learned' but their exact tensor shapes and how they interact with the lifted feature dimension are not stated explicitly, complicating re-implementation.","section":"Method section"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on our manuscript. The points raised regarding experimental controls and analysis of the gating mechanism are valid and will be addressed through revisions to strengthen attribution of results and mechanistic understanding.","responses":[{"response":"We agree that the absence of ablations and statistical reporting limits confident attribution of the gains. In the revised manuscript we will add ablation experiments with the wavelet branch disabled (alpha_wav fixed at zero) and report results over multiple random seeds including means and standard deviations.","revision_made":"yes","referee_comment":"[Experiments section] Experiments section (and abstract claim of outperformance): the manuscript states that identical hyperparameters, training data, and evaluation protocols are used, yet supplies neither ablation results (e.g., alpha_wav fixed at 0 or wavelet branch removed) nor multi-seed statistics, error bars, or variance estimates. Without these controls the observed deltas cannot be confidently attributed to the wavelet-Laplace fusion rather than optimizer stochasticity, initialization, or minor implementation differences."},{"response":"We acknowledge that further analysis of alpha_wav is needed to support the adaptive balancing claim. The revision will report converged alpha_wav values per benchmark, include sensitivity experiments to different initializations, and examine correlations between alpha_wav and performance gains on multi-scale problems such as Burgers and Navier-Stokes.","revision_made":"yes","referee_comment":"[Method section] Method section (description of fusion): the claim that the model 'adaptively balances' the two branches throughout training rests on the learned alpha_wav, but no analysis is provided of its converged values, sensitivity to initialization, or correlation with performance gains on the multi-scale problems (Burgers, Navier-Stokes). This leaves the mechanistic explanation for the reported improvements under-specified."}],"tokens_in":1547,"tokens_out":393,"duration_ms":31998,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The new piece is the concrete fusion: a parallel Haar DWT path that splits the feature map into LL/LH/HL/HH subbands, runs independent 1x1 convolutions on each, reconstructs, and mixes it into the LNO pole-residue path via a learnable sigmoid gate alpha_wav. That specific combination is not in the cited LNO work.\n\nThe architecture itself is a reasonable attempt to give LNO an explicit multi-scale spatial handle without throwing away the Laplace-domain modeling. The gate initialization that starts the wavelet contribution small is a sensible design choice.\n\nThe problem is the evidence. The abstract asserts consistent wins over LNO on five PDEs, strongest on Burgers and Navier-Stokes, yet contains zero numbers, zero error bars, zero training curves, and zero ablations. The stress-test note is correct: without a run that zeros the gate or removes the wavelet branch, and without multi-seed statistics, the deltas cannot be attributed to the added branch rather than optimizer noise or unstated implementation differences. The paper states identical protocols were used, but does not demonstrate it.\n\nThis is for people already working on neural operators who want to try a wavelet-plus-Laplace hybrid on multi-scale PDEs. A reader could pull the architectural idea and test it themselves, but the current results section does not yet support strong claims.\n\nI would send it to peer review only if the authors add the missing controls and numbers; as written the empirical core is too thin to evaluate.","headline":"WLNO adds a gated single-level Haar wavelet branch to LNO but supplies no ablations, variance numbers, or raw results to back the outperformance claim.","tokens_in":2426,"tokens_out":386,"would_cite":false,"duration_ms":16541,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Fusing a Haar wavelet branch with the Laplace Neural Operator improves performance on PDE problems with multi-scale features.","keywords":["neural operators","wavelets","partial differential equations","machine learning","Laplace transform","multi-scale decomposition","Haar wavelet","PDE solving"],"falsifier":"Running the exact same LNO and WLNO models multiple times with reported random seeds and checking if the performance difference remains significant and consistent across runs.","tokens_in":2722,"feed_emoji":"","tokens_out":607,"duration_ms":39249,"temperature":0.7,"pith_summary":"The paper introduces WLNO as an extension to LNO that incorporates a parallel Haar wavelet transform branch to capture spatially localized multi-scale features in PDE solutions. This branch decomposes the feature map into four subbands and processes them with separate convolutions before inverse transform and fusion via a learnable gate. Evaluation on five standard PDE benchmarks shows consistent gains over LNO, particularly on problems like Burgers equation and Navier-Stokes with sharp or vortical structures. A reader would care if they work with PDEs where scale separation matters for accurate prediction of dynamics.","feed_headline":"Wavelet branch lifts accuracy of Laplace neural operator on PDEs","feed_subtitle":"The hybrid model processes spatial details separately and fuses them with Laplace dynamics, showing biggest gains on shocks and vortices.","key_machinery":"The parallel wavelet branch using Haar DWT and inverse DWT with learned convolutions per subband, adaptively weighted against the Laplace branch.","core_discovery":"The central discovery is that adding a single-level Haar DWT branch with independent 1x1 convolutions on subbands, fused by a sigmoid-gated weight initialized small, to the LNO's pole-residue formulation leads to better operator learning for transient and steady-state PDE dynamics, especially where spatial multi-scale structure is prominent.","pith_inferences":["Similar wavelet augmentations might enhance other neural operator architectures beyond LNO.","Applying this to time-dependent problems with evolving multi-scale features could be a natural next test.","The approach hints at benefits from combining different frequency decompositions in operator learning."],"forward_implications":["WLNO achieves higher accuracy than LNO on the same training data and hyperparameters for diffusion, Burgers, reaction-diffusion, Darcy flow, and 2D Navier-Stokes equations.","The gains are largest for problems with sharp shock fronts and coherent vortical structures.","The learnable gate allows the model to balance the contribution of the wavelet branch during training.","The method provides an explicit way to extract multi-scale spatial features missing in pure Laplace-domain approaches."],"fun_headline_variants":["WLNO adds Haar DWT branch to LNO for PDE operator learning","Wavelet Laplace fusion outperforms LNO on five PDE benchmarks","Single level Haar wavelet augments Laplace neural operator accuracy","Gated wavelet branch complements LNO pole residue formulation"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That the observed improvements result from the wavelet branch and not from implementation variations or training differences despite claims of identical protocols.","fun_headline_variants_meta":{"raw":{"variants":["WLNO adds Haar DWT branch to LNO for PDE operator learning","Wavelet Laplace fusion outperforms LNO on five PDE benchmarks","Single level Haar wavelet augments Laplace neural operator accuracy","Gated wavelet branch complements LNO pole residue formulation"]},"model":"grok-4.3","cost_usd":0.004779,"raw_usage":{"total_tokens":2393,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":47787000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1579,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":66,"duration_ms":18461,"temperature":1.0,"reasoning_tokens":1579,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T14:54:31.613508+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the exact same LNO and WLNO models multiple times with reported random seeds and checking if the performance difference remains significant and consistent across runs.","supporting_citations":[],"review_version":1}