{"id":"0b024ba4-b4dc-4389-941b-963a480a3c4b","arxiv_id":"2507.19514","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"The paper claims that a fully spectral wavelet-domain model can reach near-Transformer accuracy on GLUE tasks while using 72% fewer parameters and no attention or convolution layers.","lead":"This paper describes a learning system that replaces neural network layers with learnable operations applied directly to wavelet coefficients, and it reports near-Transformer accuracy on sentiment classification with far fewer parameters. A generalist reader might care because it claims a route to efficient, interpretable NLP and vision models without standard neural layers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GLUE results depend on an unspecified 3D token layout; no evidence shows the reshape preserves word order, and the model's fixed local wavelet filters plus a linear head cannot be assumed to supply positional information.","rationale":"The reader rejected the paper on reproducibility grounds; I agree but want a sharper technical anchor. The claim is not merely 'no code provided,' it is that a fixed spectral pipeline on a 3D reshape of pretrained embeddings can rival Transformers. The load-bearing step is the input representation. Without a specified layout, the numbers cannot be reproduced, and there is no basis for asserting that sequential information survives. The proposed permutation check directly tests whether the result uses token order at all. If it does not, the headline claim reduces to a much weaker statement about a linear classifier on fixed wavelet features; if it does, the paper must specify the layout and show robustness across layouts. I see no independent support for this step: no code, no detailed protocol, no ablations, and no analysis of how the reshape interacts with the wavelet transform. Therefore the reader's REJECT verdict remains appropriate, and the concern I identify is the specific technical weakness that justifies it.","tokens_in":15646,"tokens_out":4778,"duration_ms":63346,"concrete_test":"Reproduce Section 5's SST-2 experiment with the exact (unspecified) pipeline, then permute the token order within each sentence before reshaping to (C,H,W), keeping embeddings, seeds, and the classifier fixed. If accuracy stays near 90%, the model is a bag-of-words-style feature extractor and the comparison to a Transformer is misleading; if accuracy drops, the result depends on an arbitrary layout that the paper must specify, justify, and validate across layouts before the central claim can be assessed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 5, Table 1: 90.1% SST-2, 84.2% MRPC, 88.5% QNLI with 0.7M parameters) depends on a text representation that the paper never specifies. Section 5 says each token sequence is 'reshaped into a three-dimensional tensor of shape (C,H,W)' with 'H × W determined by reshaping or padding,' and Section 3.2 applies a 3D DWT along axes (0,1,2), followed by soft-thresholding, gain-phase modulation, basis selection, and reconstruction. There is no positional encoding, and the only token-to-token interactions are the fixed, local wavelet filters and the final global pooling plus linear classifier. A row-major reshape into an H×W grid imposes an arbitrary 2D neighborhood that breaks adjacency at row boundaries; for MRPC and QNLI, the paper does not even state how a sentence pair is packed into one tensor. If token order is not preserved by this layout, or if the wavelet filters cannot exploit the layout, the reported accuracies are unsupported. The paper provides no layout diagram, no padding/truncation rule, no ablation over layouts, and no code, so this load-bearing assumption is entirely untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Wavelet Logic Machines,' a spectral learning framework that replaces convolutions, attention, and MLPs with a pipeline of 3D discrete wavelet transforms, learnable soft-thresholding, gain-phase modulation, and differentiable wavelet-basis selection. The main empirical claim is that this ~0.7M-parameter model reaches near-Transformer accuracy on GLUE tasks (89.3% or 90.1% on SST-2 depending on the passage) while using far fewer parameters and less memory. The paper also argues for interpretability and sketches extensions for symbolic reasoning, including spectral rule composition, cascades, a DSL, and keyed memory access.","tokens_in":15910,"tokens_out":4177,"duration_ms":48748,"significance":"If substantiated, the central claim would be significant: a parameter-efficient model without attention or convolution that approaches BERT-base on SST-2, MRPC, and QNLI would challenge standard assumptions about what language classification requires. The differentiable basis-selection mechanism and the linear-time complexity argument are genuinely interesting components. However, the paper currently provides no reproducible evidence for the GLUE results, and the empirical core is internally inconsistent. The strengths of the proposal cannot compensate for the absence of a verifiable experimental record.","major_comments":[{"comment":"The headline numbers are internally inconsistent. The abstract reports 89.3% SST-2 accuracy against a 4-layer Transformer baseline of 90.1%, with 72% fewer parameters and 58% less peak memory. Table 1 instead reports the Wavelet-Spectral model at 90.1% SST-2, lists BERT-base (93.5%) rather than any 4-layer Transformer, and Section 5 later claims peak memory reduction 'over 70%.' These are not minor wording differences; the paper cannot support two different flagship accuracy numbers and two different memory-savings figures for the same experiment. The authors must reconcile the abstract, the results table, and the efficiency text.","section":"Abstract and Table 1"},{"comment":"The language representation is not specified at a load-bearing level. The text says each token sequence is 'reshaped into a three-dimensional tensor of shape (C, H, W)' with 'H × W determined by reshaping or padding,' but there is no layout diagram, no padding or truncation rule, no description of how sentence pairs are packed for MRPC and QNLI, and no positional encoding. Because the only token-to-token interactions are fixed local wavelet filters followed by global pooling and a linear head, it is not established that word order and phrase structure survive this representation. Without a layout specification, an ablation over layouts, or code, the reported SST-2, MRPC, and QNLI accuracies are unsupported.","section":"Section 5"},{"comment":"The classification forward pass is undefined. The model in Section 3 is a denoiser: it applies DWT, soft-thresholding, gain-phase modulation, basis selection, IDWT, and minimizes reconstruction MSE. Section 5 states that the final output is pooled and passed to a single linear layer with a cross-entropy loss, but no equation or architecture for this classification path is given, and there is no explanation of how the denoising parameters (λA, λD, γ, θ, α) relate to the classifier. This makes the GLUE results impossible to check against the model definition.","section":"Sections 3.2 and 5"},{"comment":"The experimental protocol is missing. The paper gives no dataset splits, tokenizer, embedding dimension, sequence length, batch size, learning rate, number of epochs, or number of runs for any GLUE result. The baselines are said to be trained 'under identical conditions,' but the BERT-base fine-tuning details are absent, and no code is provided. Given that a 0.7M-parameter model without contextual layers is claimed to outperform a 2.1M-parameter BiLSTM on these benchmarks, the lack of reproducibility evidence is a load-bearing gap.","section":"Section 5 and Table 1"},{"comment":"The interpretability and reasoning claims are not backed by the experiments. Section 6 asserts that in GLUE experiments the model assigns high probability to Daubechies-4 and Symlet-6 bases, and Section 7 claims that Level 1 might isolate negations, but no quantitative analysis, learned-weight inspection, or visualization from the actual GLUE runs is provided. Section 8 analyzes synthetic sentence embeddings with the Fourier transform, not the trained model's wavelet representations. These sections therefore do not support the paper's advertised interpretability and reasoning contributions.","section":"Sections 6, 7, and 8"}],"minor_comments":[{"comment":"There are typographical errors: 'W avelet-Spectral' and 'BER T-base' should be 'Wavelet-Spectral' and 'BERT-base.'","section":"Table 1"},{"comment":"The reference list contains duplicates (e.g., [23]/[8], [24]/[9], [26]/[11], [34]/[20]) and inconsistent numbering; please deduplicate and renumber.","section":"References"},{"comment":"The text says the model 'applies 2D or 3D DWT' while the rest of the paper claims full 3D support; the GLUE experiments should specify which transform is actually used.","section":"Section 5"},{"comment":"The phrase 'Without Neural Networks' is not consistent with the final linear classifier and pretrained embeddings; the paper should clarify that it eliminates hidden neural layers, not all learned linear components.","section":"Title and Abstract"},{"comment":"The five reasoning modules are speculative and are not evaluated on any task; consider moving this material to future work or adding experiments.","section":"Section 9"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft with several unreconciled numerical claims and an underspecified experimental core. The self-citations [15-17] do not provide external validation. If the authors can supply code, a precise GLUE protocol, and a corrected results table, a revised submission could be reconsidered, but in its current form the central empirical claim is not verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the wavelet logic machine paper. The core idea has some genuine content: combining learnable soft-thresholding, gain/phase modulation, and softmax basis mixing over wavelet coefficients—with no convolution, attention, or MLP layers—is a real extension of classical wavelet shrinkage and of spectral transformer work, including the author's own earlier papers. The wavelet mathematics in Sections 2–4 is standard, clearly written, and the architecture is coherent as a denoiser. That part deserves credit.\n\nThe empirical core does not hold up. The abstract reports 89.3% SST-2 accuracy against a 4-layer Transformer baseline at 90.1%; Table 1 reports 90.1% for the wavelet model and BERT-base at 93.5%, with no 4-layer Transformer appearing anywhere. Memory savings are 58% in the abstract and \"over 70%\" in the body. Those are not rounding differences. There is also no code, no data, no seeds, no hyperparameters, and no evaluation protocol, so none of the headline numbers can be checked. The stress-test note about the token layout is on target: Section 5 says each sequence is reshaped into a (C,H,W) tensor, but the paper never specifies the layout, the padding rule, or how sentence pairs are packed into one tensor. The model has no positional encodings, and fixed local wavelet filters plus global pooling cannot silently supply word order. That load-bearing assumption is entirely untested. Sections 9–11 add reasoning modules and Laplacian-wavelet extensions with no implementation at all—speculative filler rather than evidence.\n\nWhere does this leave things? If someone reimplemented this and the numbers held with consistent reporting, the idea would be a reasonable efficient alternative for low-resource NLP and signal processing. But as submitted, the paper is not reproducible and contradicts itself on its main results. I would not cite it, and I would not send it to a serious referee in this state. The right move is desk reject, ideally with an invitation to resubmit with code and corrected, internally consistent experiments—or with the GLUE claims dropped until they are supported. This is a paper to watch, not one to trust yet.","headline":"A genuinely interesting spectral-learning architecture, but the GLUE results are internally inconsistent, unreproducible, and rest on an unspecified token layout; desk reject rather than referee.","tokens_in":16458,"tokens_out":2825,"would_cite":false,"duration_ms":34739,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a model with no convolution, attention, or MLP layers—only learnable wavelet-domain operations—achieves near-Transformer accuracy on SST-2 with 72 percent fewer parameters.","keywords":["wavelet transform","spectral learning","soft-thresholding","learnable basis selection","GLUE benchmark","SST-2 sentiment classification","model efficiency","interpretability"],"falsifier":"Re-run the reported GLUE experiments from the recipe in Section 5: fixed pretrained token embeddings, 3D reshape, wavelet soft-thresholding, and no positional encodings. If SST-2 accuracy does not land near 90 percent, or if the 0.7M-parameter count and the claimed memory savings do not reproduce, the paper's central quantitative claim is unsupported.","tokens_in":15377,"feed_emoji":"🌊","tokens_out":9279,"duration_ms":96436,"temperature":0.7,"pith_summary":"The paper tries to show that learning does not require neural network layers: a model that works only on wavelet coefficients, with learnable shrinkage, gain-phase modulation, and a softmax choice among wavelet bases, can denoise 3D signals and classify natural-language sentences. On GLUE tasks the model reports accuracy near 90 percent on SST-2, close to a four-layer Transformer (90.1 percent) while using 72 percent fewer parameters and 58 percent less peak memory. The authors claim faster early convergence from the spectral sparsity prior and $O(n)$ instead of $O(n^2)$ complexity for long inputs. If this holds, a compact, inspectable, spectral model is a viable alternative to overparameterized neural architectures for at least some vision and language workloads.","feed_headline":"Wavelet-only model nears 90% on SST-2 sentiment","feed_subtitle":"No attention, no MLPs: learnable thresholding and basis selection do the work with a fraction of the parameters.","key_machinery":"The load-bearing object is the 3D discrete wavelet transform (DWT) and its inverse, used as the only feature-extraction step. Around it sit three learnable mechanisms: the pointwise spectral nonlinearity $\\phi(z) = \\gamma\\, \\mathrm{sign}(z)\\,\\max(|z|-\\lambda,0)\\,\\cos\\theta$ with learnable threshold $\\lambda$, gain $\\gamma$, and phase $\\theta$; a softmax over learnable logits $\\alpha_k$ that mixes reconstruction paths from candidate wavelet families (Haar, Daubechies, Symlet, and Biorthogonal); and an entropy penalty that prunes low-weight bases. The argument is that these operations, applied subband by subband, supply the filtering and feature selection that neural networks normally learn with spatial filters.","core_discovery":"On the paper's own terms, the discovery is that a complete learning pipeline can be built from wavelet-domain operations alone: the input is decomposed by a discrete wavelet transform, the coefficients are modified by learnable soft-thresholding and gain-phase modulation, and the inverse transform reconstructs a representation that is either a denoised signal or the basis for a classifier. The model has no convolution, attention, or MLP layers, and no positional encodings; the only trainable quantities are thresholds $\\lambda$, gains $\\gamma$, phase shifts $\\theta$, and softmax basis-selection logits $\\alpha_k$. The author reports that this 0.7M-parameter spectral model reaches roughly 90 percent accuracy on SST-2 and competitive scores on MRPC and QNLI, close to a four-layer Transformer, using 72 percent fewer parameters and 58 percent less peak memory. The intended lesson is that spectral sparsity and multiscale structure carry much of the inductive bias that deep neural architectures are normally assumed to provide.","pith_inferences":["The paper's Section 9 reasoning extensions are presented as proposals without experiments; a concrete next step is running them on a small reasoning benchmark to see whether spectral cascades or keyed memory improve accuracy beyond the base pipeline.","If the 3D reshape is truly order-insensitive, the SST-2 result would suggest sentiment classification is largely decodable from local frequency statistics of token embeddings; shuffling token order before reshaping would test this directly.","The abstract reports 89.3 percent while Table 1 reports 90.1 percent for the same model, and memory savings are given as both 58 percent and 'over 70 percent' in different passages, so an exact reproducible run is needed to pin down the quantitative claim.","Because the language pipeline starts from fixed pretrained embeddings, part of the result may be inherited from those embeddings; replacing them with learned or contextual embeddings would isolate the contribution of the wavelet mechanism."],"forward_implications":["A linear-time wavelet pipeline would remove the $O(n^2)$ attention bottleneck for long sequences, lowering inference cost on large inputs.","If the GLUE numbers hold, a 0.7M-parameter model would come within a few points of the large pretrained transformer baseline on SST-2, MRPC, and QNLI while being over 150x smaller.","Learned thresholds and gains give a per-subband record of what the model suppresses or amplifies, enabling frequency-localized explanations for predictions.","The differentiable basis selector lets the model choose different wavelet families per task, adapting its inductive bias without manual filter design.","The spectral sparsity prior could speed early training convergence, reducing compute before full convergence."],"supporting_citations":[{"why":"Defines the multiresolution wavelet decomposition that the whole model operates on.","marker":"[27]"},{"why":"Supplies the soft-thresholding shrinkage operation that the model makes learnable.","marker":"[30]"},{"why":"Defines the GLUE benchmark and the SST-2, MRPC, and QNLI tasks used for evaluation.","marker":"[26]"},{"why":"The large pretrained transformer baseline whose accuracy the spectral model is claimed to approach.","marker":"[19]"},{"why":"Provides the self-attention architecture and the quadratic-complexity baseline being replaced.","marker":"[34]"},{"why":"Supplies the fixed pretrained token embeddings that are reshaped into 3D tensors.","marker":"[18]"},{"why":"Provides the discrete wavelet transform implementation used in the experiments.","marker":"[40]"}],"fun_headline_variants":["Wavelet-only classifier hits 89.3% on sentiment, no neural nets","Spectral learning without neural networks: 89.3% on SST-2","No attention, no MLPs: wavelet learning nears Transformer accuracy","Wavelet transform replaces neural layers, nears 90% on SST-2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reshaping a token sequence into a 3D tensor and processing it with wavelets, without any positional encoding, preserves the word-order and phrase-level information needed for sentiment and inference classification.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet-only classifier hits 89.3% on sentiment, no neural nets","Spectral learning without neural networks: 89.3% on SST-2","No attention, no MLPs: wavelet learning nears Transformer accuracy","Wavelet transform replaces neural layers, nears 90% on SST-2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3340,"prompt_tokens":987,"completion_tokens":2353,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2269}},"tokens_in":603,"tokens_out":2353,"duration_ms":21178,"temperature":1.0,"reasoning_tokens":2269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:21:01.597677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the reported GLUE experiments from the recipe in Section 5: fixed pretrained token embeddings, 3D reshape, wavelet soft-thresholding, and no positional encodings. If SST-2 accuracy does not land near 90 percent, or if the 0.7M-parameter count and the claimed memory savings do not reproduce, the paper's central quantitative claim is unsupported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the multiresolution wavelet decomposition that the whole model operates on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the soft-thresholding shrinkage operation that the model makes learnable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the GLUE benchmark and the SST-2, MRPC, and QNLI tasks used for evaluation."},{"cited_title":"Devlin, M","cited_arxiv_id":null,"evidence_quote":"The large pretrained transformer baseline whose accuracy the spectral model is claimed to approach."},{"cited_title":"Pennington, R","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed pretrained token embeddings that are reshaped into 3D tensors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the discrete wavelet transform implementation used in the experiments."}],"review_version":1}