{"id":"18f70553-2a93-4a26-8259-ddf376f43a31","arxiv_id":"2508.00420","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A non-parameterized sentence embedding method that first compresses word vectors with wavelets and then applies a cosine transform over the compressed coefficients, matching baseline performance at a fraction of the dimension.","lead":"This paper applies discrete wavelet transforms to word embeddings to compress them, then combines wavelets with discrete cosine transforms to build fixed-size sentence embeddings. The authors report that these compressed embeddings match or beat standard averaging and DCT baselines on several language tasks while using far smaller vectors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported advantage of DWT-DCT over DCT-only is not yet established because the paper allows per-task selection of wavelet family, level, and coefficients without a validation protocol, and the numeric gains are smaller than evaluation noise.","rationale":"The reader identified the local-correlation assumption as the weakest point; I agree that the paper does not prove correlation between neighboring embedding dimensions. However, the empirical evidence, especially the random-pooling baseline in Table 6, already suggests DWT is not equivalent to random compression. The more load-bearing weakness is methodological: the paper's own Section 6.1 reveals per-task selection of wavelets and scales, and the headline gains over DCT are small and unreplicated. This is a correctness risk rather than a novelty objection, and the fixed-configuration test would settle whether the reported superiority survives. Because the reader's CONDITIONAL verdict already accounts for this uncertainty, no verdict change is needed.","tokens_in":12763,"tokens_out":4491,"duration_ms":47317,"concrete_test":"Use code if released, or re-implement, and fix one configuration: Level-1 DWT with Coiflet1, K=1, skip trick disabled, MLP settings from [2]. Run all tasks in Table 7 with this configuration and with DCT c[0] under identical scripts, reporting mean and standard deviation over at least 5 random seeds or bootstrap resamples. If the fixed-configuration DWT-DCT does not beat DCT c[0] on a majority of tasks, the claimed advantage is attributable to per-task model selection rather than to DWT.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Table 7 rests on a configuration that is not fixed. Section 6.1 says 'we simplified our paper by omitting the specific details of MW and its application in each task,' and Table 6's caption indicates the mother wavelet used per condition per task is shown in italics. Thus the reported numbers may be the best of several per-task choices (wavelet family, scale, level, K, and the skip trick) rather than the output of a single non-parameterized method. The headline differences are small: DWT1-DCT[0] beats DCT c[0] by 1.33 points on CR, 1.15 on SST2, and 2.16 on SST5, while being 0.07 worse on MR; no error bars or significance tests are reported. Per-task selection on the test set could easily account for differences of this size. In addition, the DCT baseline numbers are taken from [2] rather than re-run in the same pipeline, so implementation differences are uncontrolled. Until a fixed configuration and a proper validation protocol are reported, the claim of 'comparable and even superior' results is not fully supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using the Discrete Wavelet Transform (DWT) to compress word embeddings along their dimension axis, and combining DWT with the Discrete Cosine Transform (DCT) to produce fixed-size sentence embeddings. The method is deterministic and non-parameterized: it applies a mother wavelet row-wise to the sentence embedding matrix, then applies DCT column-wise over the resulting coefficient matrices, keeping one or two DCT coefficients. The authors evaluate the word-level embeddings on similarity and categorization tasks, and the sentence-level embeddings on SentEval classification and semantic textual similarity tasks, comparing against averaging, DCT, EigenSent, p-means, and VLAME baselines. They report that DWT-DCT matches or improves on DCT-only sentence embeddings at substantially lower dimensionality.","tokens_in":13041,"tokens_out":6234,"duration_ms":59276,"significance":"The core idea is simple, novel in combination, and, if validated, practically useful: a fixed-size sentence embedding that is cheap, deterministic, and does not require training. The paper's breadth of evaluation (word similarity, categorization, and nine downstream tasks) and the qualitative analyses are assets. However, the manuscript currently does not supply a reproducible configuration (the mother wavelet and scale are selected per task without a validation protocol), reports no variance or significance measures, and relies on external baseline numbers for the central comparison. These issues must be resolved before the central claim of 'comparable and even superior' performance can be accepted.","major_comments":[{"comment":"The central comparison in Table 7 is not based on a fixed configuration. Section 6.1 states that 'the specific details of MW and its application in each task' are omitted, and Table 6's caption indicates that the mother wavelet used per condition per task is shown in italics. Because the method is presented as non-parameterized (Section 3), the per-task choice of wavelet family, scale, number of DWT levels L, and the K=2 skip trick are effectively free choices made after seeing the results. Without a pre-registered or nested validation protocol that fixes these choices on development data only, the reported gains (e.g., CR 81.14 vs 79.81 in Table 7) may reflect test-set selection. Please report the exact wavelet configuration per task and a validation-based selection procedure.","section":"§6.1, Tables 6 and 7"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the tables. The differences that support the headline claims are small: in Table 7, DWT1-DCT[0] beats DCT c[0] by 1.33 points on CR and 1.15 on SST2, while being 0.07 points worse on MR. Since SentEval classifiers are stochastic, these gaps are likely within run-to-run noise. In addition, the DCT baseline rows in Table 7 are taken from [2] ('as reported in [2]') rather than re-run in the same codebase, so differences in tokenization, classifier hyperparameters, or evaluation versions are uncontrolled. Please re-run all baselines in the same pipeline and report variance or bootstrap confidence intervals.","section":"§4.3 and §5.1, Tables 6 and 7"},{"comment":"Table 6 contains corrupted entries that prevent interpretation. The row A VG(cD) contains the value '76.61coif15' in the MR column, and the A VG(Random Pooling) row contains '8' for TREC and '33' for STS12, which are implausible accuracy and Pearson correlation values. These entries must be corrected or explained, and the table caption should be checked for the repeated model name 'A VG(cD+cAD+cAAD)'.","section":"Table 6"},{"comment":"The paper admits that 'we did not explicitly prove that the dimensions in the same embedding are correlated.' This correlation is the premise for applying a wavelet filter along the index order of embedding dimensions. The argument that DWT achieves comparable results therefore presupposes the conclusion. Please add a direct test, for example comparing DWT compression to the same transform applied to randomly permuted embedding dimensions and to a random projection of the same output dimension. Without such a control, the proposed mechanism for why DWT helps is not established.","section":"§6.2"}],"minor_comments":[{"comment":"The description of the DWT coefficient matrices is confusing: for L=2, the text lists cDD, cDA, cAD, and cAA, but Section 4.2 uses 'cD+cAD' and 'cA+cDA' without defining how these matrices are concatenated. Please clarify the naming and the concatenation order.","section":"§3"},{"comment":"The conclusion refers to 'DCT[0:2]' but Table 7 only shows DCT c[0:1]. Please align the notation between the text and the table.","section":"§6"},{"comment":"The text states that averaging (AVG) with FastText embeddings is the baseline, but Table 6 also lists 'A VG(Random Pooling)' and 'A VG(PCA+PPA)'. Please clarify whether all baselines use FastText and how the random pooling condition was constructed.","section":"§4.3"},{"comment":"The captions say that the best results are shown in red, but color is not visible in a printed or grayscale copy. Please use bold or another typographic distinction that survives monochrome printing.","section":"Tables 1, 6, 7"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection. The core idea is plausible and the paper is not fatally flawed, but the evaluation needs to be made honest and reproducible. The most important issue is the per-task selection of wavelet family, scale, and levels without a validation protocol; if the authors fix the configuration a priori or report a proper validation-based selection, the central claim may survive. The missing correlation proof is a concern but can be addressed with a permutation control. I would also ask the authors to re-run the DCT baselines in the same codebase and report variance, as the current comparisons are not yet convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is narrow but real. It applies a discrete wavelet transform to word embeddings, then combines the compressed word vectors with the DCT sentence embedding from Almarwani et al. [2] to get fixed-size sentence representations. That combination is new relative to the cited DCT-only work, and the evaluation is broad: intrinsic word similarity, concept categorization, and a battery of SentEval downstream tasks. The method is genuinely non-parameterized, and the linear complexity argument (O(LN)) is straightforward and correct.\n\nThe soft spot is the evaluation protocol. Section 6.1 says the authors 'simplified our paper by omitting the specific details of MW and its application in each task,' and Table 6's caption says the mother wavelet used per condition per task is shown in italics. That means the headline Table 7 numbers may be the best of several per-task choices (wavelet family, level, K, and the skip trick) rather than the output of a single fixed method. The gains over DCT c[0] are small: 1.33 points on CR, 1.15 on SST2, 2.16 on SST5, and -0.07 on MR. With no error bars and no significance tests, those differences sit inside ordinary evaluation noise. The DCT baseline numbers come from [2] instead of being re-run in the same pipeline, so implementation differences are uncontrolled. The authors need to fix a configuration, describe a proper validation protocol, and report variance.\n\nThere are also presentation problems: a table entry reads '76.61coif15' — the wavelet name glued onto the score — which points to sloppy proofreading. The assumption that word-embedding dimensions are locally correlated along the index is asserted, not proven; the authors admit this in Section 6.2. That admission is honest, but it means the mechanism behind the method is still a conjecture.\n\nOn balance this is a plausible incremental idea with broad but under-controlled evidence. I'd send it to peer review rather than desk reject, because the idea is testable and the authors are transparent about their own limitations. The referee's main job should be to force a fixed configuration, variance reporting, and re-run DCT in the same setup. I would not lean on these results in my own work yet, but once cleaned up the paper would be citable as prior art for DWT-DCT sentence embeddings.","headline":"A plausible, non-parameterized DWT-DCT sentence embedding that needs a fixed configuration and variance reporting before its headline gains over DCT-only can be trusted.","tokens_in":13546,"tokens_out":2382,"would_cite":true,"duration_ms":22562,"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":"The paper proposes a parameter-free combination of discrete wavelet and cosine transforms that encodes sentences into fixed-size vectors of the word-embedding dimension, matching or exceeding larger baselines on several NLP tasks.","keywords":["sentence embedding","discrete wavelet transform","discrete cosine transform","word embedding compression","parameter-free","spectral methods","non-parametric sentence embedding","natural language processing"],"falsifier":"Permute the coordinate order of every word embedding with one shared random permutation and rerun the reported experiments. If downstream accuracy stays close to the unpermuted case, the wavelet stage is not exploiting the local ordering of dimensions; if accuracy collapses to random-pooling levels, the local-correlation assumption is the load-bearing premise.","tokens_in":12582,"feed_emoji":"🧠","tokens_out":10069,"duration_ms":92415,"temperature":0.7,"pith_summary":"The paper claims that discrete wavelet transforms (DWT) and discrete cosine transforms (DCT) can be applied together to turn variable-length sentences into fixed-size vectors without losing task performance. If true, this gives a simple, parameter-free alternative to methods that either average word vectors, losing word order, or concatenate many DCT coefficients, producing very long vectors such as 1800 dimensions for six coefficients on 300-dimensional words. The proposal applies DWT along each word vector to compress it, then DCT across the words in the sentence to encode word order, ending with a vector the size of one word embedding. Experiments on sentiment, paraphrase, and question-type tasks show results comparable to or better than averaging and DCT-only baselines, often at a quarter or half the dimension. The paper also presents evidence that DWT coefficients alone can compress word embeddings to 50–75% of their original size while preserving semantic similarity.","feed_headline":"Wavelet-cosine embeddings pack sentences into word-sized vectors","feed_subtitle":"A parameter-free transform pair compresses sentences to word size, beating DCT-only and averaging baselines.","key_machinery":"The central object is the conjoint DWT-DCT transform, which first applies a discrete wavelet transform row-wise to the sentence matrix (each row is a word vector), splitting every word into approximation and detail coefficient vectors at one or more levels, and then applies a discrete cosine transform column-wise across the words of the sentence; the final sentence vector is formed by concatenating one or two DCT coefficients from each coefficient matrix, with a skipping trick that keeps the total dimension equal to the word-embedding dimension $d$ instead of $Kd$. This two-stage pipeline is what lets the method compress both the intra-word structure (via DWT) and the inter-word order (via DCT) into a fixed-size vector.","core_discovery":"The central discovery is that word embeddings behave enough like signals with locally correlated dimensions that a fixed wavelet low-pass/high-pass split can separate salient semantic content from detail, and the resulting coefficients can be summarized across a sentence with DCT without needing learned parameters. The authors show that Level-1 approximation (cA) and detail (cD) coefficients, at half the dimension, match or exceed original embeddings on word-similarity tasks, and that combining them with DCT produces sentence vectors of dimension d (the word-embedding size) that outperform DCT-only vectors of dimension Kd and even match or beat averaging on most tasks. They also report qualitative evidence that low-pass coefficients capture category-level commonalities such as animals or emotions, while high-pass coefficients preserve contrasts such as gender or age.","pith_inferences":["The paper does not compare against learned pooling operators of the same size; a natural next test would be to replace the fixed wavelet filters with a small learned 1-D convolution, which would show whether the fixed basis is near-optimal or leaves performance on the table.","Because the method fixes the sentence vector to the word dimension, it could be used as a cheap input projection for long-document or sequence models, potentially reducing the cost of self-attention over many tokens.","The qualitative split between approximation coefficients (which capture category-level similarity) and detail coefficients (which preserve contrasts) suggests that a task-specific weighting of the two channels could improve results further, an option the paper does not explore."],"forward_implications":["Sentence embedding size stays at the word-embedding dimension $d$ instead of growing as $Kd$, where $K$ is the number of DCT coefficients, so downstream memory and computation drop sharply for the same information.","Because the transform is parameter-free, it applies directly to any pretrained word embeddings without training, making it a drop-in compression for existing pipelines.","DWT alone compresses word embeddings to 50–75% of their original size while preserving word-similarity and categorization performance, which is useful when storage or retrieval speed is the bottleneck.","On several tasks (CR, SST2, MRPC, STS12), the DWT-DCT model outperforms both averaging and DCT-only baselines, indicating that the two transforms capture complementary information rather than redundant signal."],"supporting_citations":[{"why":"Supplies the DCT sentence-embedding baseline and the K-coefficient concatenation scheme that the proposed DWT-DCT method compresses and extends.","marker":"[2]"},{"why":"Provides the standard evaluation tasks and MLP classifier setup used for all extrinsic sentence-embedding comparisons.","marker":"[11]"},{"why":"Supplies the EigenSent spectral sentence-embedding baseline against which the proposed model is compared at the same dimension.","marker":"[23]"},{"why":"Provides one set of pretrained word embeddings used in both intrinsic and extrinsic evaluations.","marker":"[28]"},{"why":"Provides the other set of pretrained word embeddings used for word-similarity and concept-categorization experiments.","marker":"[34]"},{"why":"Supplies the PCA-based dimensionality-reduction baseline (with post-processing) that DWT word embeddings are compared against.","marker":"[35]"},{"why":"Supplies the P-Means sentence-embedding baseline, a non-parameterized competitor with much larger dimension.","marker":"[37]"},{"why":"Supplies the VLAWE sentence-embedding baseline, another non-parameterized competitor with much larger dimension.","marker":"[22]"}],"fun_headline_variants":["Wavelet-cosine transform packs sentences into word size","Transform pair shrinks sentences to word-sized vectors","Wavelets on word embeddings beat averaging baselines","DWT+DCT embeddings rival full-size sentence vectors","Wavelet sentence embedding: word-size, no parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the dimensions of a word embedding are locally correlated with their neighbors in the vector, so that fixed wavelet low-pass filters that average nearby coordinates preserve the semantics needed for downstream tasks; the paper states it does not prove this correlation.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet-cosine transform packs sentences into word size","Transform pair shrinks sentences to word-sized vectors","Wavelets on word embeddings beat averaging baselines","DWT+DCT embeddings rival full-size sentence vectors","Wavelet sentence embedding: word-size, no parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1171,"prompt_tokens":847,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":463,"tokens_out":324,"duration_ms":3736,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:08:09.067712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the coordinate order of every word embedding with one shared random permutation and rerun the reported experiments. If downstream accuracy stays close to the unpermuted case, the wavelet stage is not exploiting the local ordering of dimensions; if accuracy collapses to random-pooling levels, the local-correlation assumption is the load-bearing premise.","supporting_citations":[{"cited_title":"Efficient sentence embedding using dis- crete cosine transform","cited_arxiv_id":null,"evidence_quote":"Supplies the DCT sentence-embedding baseline and the K-coefficient concatenation scheme that the proposed DWT-DCT method compresses and extends."},{"cited_title":"EigenSent: Spectral sentence embeddings using higher-order dynamic mode decomposition","cited_arxiv_id":null,"evidence_quote":"Supplies the EigenSent spectral sentence-embedding baseline against which the proposed model is compared at the same dimension."},{"cited_title":"Advances in pre-training distributed word representations","cited_arxiv_id":null,"evidence_quote":"Provides one set of pretrained word embeddings used in both intrinsic and extrinsic evaluations."},{"cited_title":"GloVe: Global vectors for word representation","cited_arxiv_id":null,"evidence_quote":"Provides the other set of pretrained word embeddings used for word-similarity and concept-categorization experiments."},{"cited_title":"Effective dimensionality reduction for word embeddings","cited_arxiv_id":null,"evidence_quote":"Supplies the PCA-based dimensionality-reduction baseline (with post-processing) that DWT word embeddings are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the VLAWE sentence-embedding baseline, another non-parameterized competitor with much larger dimension."}],"review_version":1}