{"id":"28e03e44-51d1-457d-9824-e74ecb192c8b","arxiv_id":"2505.20176","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Placing a KAN layer between two linear layers improves spoken language understanding accuracy over linear-only baselines on several speech-intent datasets.","lead":"This paper tests whether Kolmogorov-Arnold Networks, a newer type of neural network layer, can replace plain linear layers in spoken language understanding models. It finds that a specific placement, one KAN layer between two linear layers, gives the best accuracy on several speech-intent datasets. The result matters because it suggests a practical architectural swap that can slightly improve speech assistants without adding much training cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'most cases' claim is not statistically established: FKF and the B-spline approximation were selected on FSC/Timers and Such, the same datasets that show the largest FKF gains, while the three untouched datasets show differences within reported noise.","rationale":"The Reader's weakest assumption states that Table 4 is reported without standard deviations; in the manuscript, Table 4 actually includes standard deviations for every entry, so that specific factual claim is wrong. The underlying worry about statistical robustness is nevertheless legitimate, but it is better located in the selection procedure. FKF was selected from Table 1 on FSC and Timers and Such, and B-splines were selected from Table 3 on FSC; Table 4 then evaluates FKF on those same datasets and counts the wins. The datasets that were not involved in selection show only small, likely non-significant differences, and on SLURP FKF is numerically worse. The paper has real strengths: it compares multiple configurations and approximation functions, uses a standard efficient-KAN implementation, and promises code. These make the central claim plausible but conditional on a sharper statistical demonstration, so the CONDITIONAL verdict stands.","tokens_in":10001,"tokens_out":6491,"duration_ms":72050,"concrete_test":"Re-run FFF and FKF on the three datasets not used for configuration selection (SLURP, ITALIC, and Speech-MASSIVE German/French) with 10 fixed random seeds and identical hyperparameters, then compute paired 95% bootstrap confidence intervals for the F1 difference FKF minus FFF on each dataset. Require the interval to exclude zero on at least three of the four dataset/language settings before accepting 'superior in most cases'; otherwise report the claim as unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FKF (one KAN layer between two linear layers) achieves comparable or superior performance in most cases and that this generalizes across architectures. The most fragile premise is that the transformer results in Table 4 are independent evidence; they are not for the two datasets that drive the effect. FKF was chosen from five configurations using 2D-CNN results on FSC and Timers and Such (Table 1), and the B-spline approximation was chosen on FSC (Table 3). The transformer evaluation in Table 4 then reuses exactly FSC and Timers and Such, and the largest FKF-vs-FFF deltas (+0.005 and +0.016 F1) occur on those selected datasets. On the three datasets that were not used in any selection step (SLURP, ITALIC, Speech-MASSIVE de-DE/fr-FR), the deltas are -0.008, +0.002, +0.008, and +0.007 F1, all within one standard deviation of the printed FFF baseline. No seed count or paired significance test is reported, so these small differences cannot be distinguished from run-to-run noise. Thus the phrase 'most cases' is not supported by statistically independent comparisons; it depends on reusing the selection datasets in the headline evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether Kolmogorov-Arnold Network (KAN) layers can replace linear layers in the classification heads of Spoken Language Understanding (SLU) models. It first searches over five FF/KAN configurations on a 2D-CNN trained on FSC and Timers and Such, identifies FKF (a KAN layer between two linear layers) as the best configuration, and ablates the KAN hidden size and approximation function. It then transfers FKF to wav2vec 2.0 and XLS-R and reports F1-macro results on five SLU datasets, followed by a qualitative analysis of word-level explanations on Timers and Such. The central empirical claim is that KAN layers can effectively replace linear layers with comparable or superior performance in most cases.","tokens_in":10348,"tokens_out":6542,"duration_ms":63173,"significance":"The paper provides a systematic configuration search for KAN placement in SLU classification heads, releases code, and includes ablations over hidden size and approximation function. If the reported gains are reproducible, the result is a modest but useful architectural substitution for SLU models. The main limitation is statistical: the headline transformer comparison reuses the datasets used for model selection, no seed count or significance testing is reported, and most cross-dataset differences are within one standard deviation of the baseline. The paper also contains an incorrect definition of the KAN layer in Eq. (1).","major_comments":[{"comment":"The claim that FKF 'generalizes well' across transformer architectures is not statistically supported by the reported numbers. FKF was selected using the 2D-CNN on FSC and Timers and Such (Tables 1–3), and Table 4 reuses those two datasets; the largest FKF-vs-FFF deltas (+0.005 and +0.016 F1) occur there. On the datasets not used in selection (SLURP, ITALIC, SPEECH-MASSIVE de-DE and fr-FR), the deltas are -0.008, +0.002, +0.008, and +0.007, all within one standard deviation of the FFF baseline. The paper neither states the number of seeds nor reports paired significance tests, so these differences cannot be distinguished from run-to-run noise. Please report seed counts and significance tests, or explicitly cast FSC and Timers and Such as selection data and evaluate the generalization claim only on the untouched datasets.","section":"§5.2, Table 4"},{"comment":"The exact KAN configuration transferred to the transformer models is not specified. Section 5.1 shows that FKF performance depends strongly on the hidden-size design (fixed vs variable) and hidden size (Table 2), with the best 2D-CNN result at FKF(b) with 512 hidden units (0.631 accuracy vs 0.555 for FFF). The model sizes in Table 4 (e.g., 95.4M vs 94.6M for wav2vec2) suggest a much smaller KAN head, but the paper never states which variant, (a) or (b), and which hidden size were used in the transformer experiments. Please specify the exact architecture and justify the choice relative to the ablation.","section":"§5.2, Table 4; §4"},{"comment":"The attention analysis is based on a single example and the statement that the 'pattern holds across other misclassified samples' is not accompanied by any quantitative aggregation, error counts, or multi-run evidence. This overstates the reliability of the explanation-level conclusion. Please provide aggregate statistics (e.g., the number of corrected cases and the proportion that follow the stated pattern) or soften the claim to an anecdotal observation.","section":"§5.3, Fig. 3"},{"comment":"All tables report standard deviations but no number of seeds or runs. The abstract's 'comparable or superior performance in most cases' is a claim about average performance; without knowing whether the reported values are single runs or averages over N seeds, the standard deviations cannot be interpreted and the claim is not falsifiable from the reported numbers. Add the number of runs to every table caption and, ideally, report paired significance tests between configurations.","section":"Tables 1–4"}],"minor_comments":[{"comment":"The displayed formula for a KAN layer appears to use products over input dimensions; a KAN layer is a sum of univariate functions. The equation should read L(x) = sum over i of phi_{i,o}(x_i), not a product.","section":"§2, Eq. (1)"},{"comment":"The text says FFK and FKF 'consistently outperform FFF in most datasets,' but on SLURP both FFK and FKF have lower F1 than FFF (0.534 and 0.531 vs 0.539). Please rephrase to reflect the actual pattern.","section":"§5.2, first paragraph"},{"comment":"The footnote markers in 'wav2vec 2.02' and 'XLS-R3' render poorly in the manuscript; use proper citation or inline references.","section":"§4, Models"},{"comment":"The notation FKF(a) and FKF(b) is introduced only in the table caption; define these two variants in the text before referring to them in the hidden-size analysis.","section":"Table 2, §5.1"},{"comment":"The phrase 'without increasing model size or training time' is not exact: FKF is 6.3M vs 6.2M in Table 1 and 95.4M vs 94.6M in Table 4. Please say 'without substantially increasing model size or training time.'","section":"Introduction, last bullet"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental but potentially useful empirical study. The central issue is statistical: the headline result reuses selection datasets and lacks seed counts and significance tests. With multi-seed results and a clearer specification of the transferred configuration, the paper could be acceptable for a workshop or conference; in its current form, the 'most cases' claim is not yet supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a modest but genuinely useful paper: the first systematic look at Kolmogorov-Arnold layers in spoken language understanding, as opposed to the keyword spotting and speech enhancement work cited. The authors try five FF/KAN configurations in a 2D-CNN dense block, pick the best (FKF, one KAN between two linear layers), then carry that configuration to wav2vec 2.0 and XLS-R across five datasets. They also ablate hidden size and approximation functions. The experiments are broad and the negative results are reported plainly: full KAN replacement hurts, placement matters, B-splines beat RBF/Chebyshev/RSWAF/GR-KAN in this setup. That is the kind of empirical information people actually use.\n\nThe soft spots are real but not fatal. FKF and the B-spline approximation were selected on FSC and Timers and Such, and those same two datasets produce the largest transformer gains (+0.005 and +0.016 F1). On the three datasets not used in selection, the deltas are within one printed standard deviation of baseline. So the \"superior\" part of \"comparable or superior\" is not statistically established on independent data. What is established is \"comparable\": FKF loses only on SLURP and is within noise elsewhere. The stress-test note is right about the selection effect. One correction to the reader's take: Table 4 does report standard deviations; the missing piece is the number of seeds and any paired significance test. Without seed counts we can't know whether the 0.001-0.008 differences are meaningful.\n\nThe abstract's \"without increasing model size\" is slightly off: FKF adds ~0.8M parameters over the baseline in both transformer settings, which is trivial but not zero. The attention analysis is a single example with a plausible story, not a systematic evaluation; the authors themselves present it as insight, so that's acceptable.\n\nOverall: a solid, honest engineering paper that would benefit from a revision reporting seeds, paired tests, and a sharper separation between selection datasets and held-out ones. The effect is small, but the question is well-posed and the config search is more thorough than most. I'd send it to a referee.","headline":"First systematic KAN-for-SLU study with real configuration search; the headline gains are partly selection effects, but the paper is honest and worth refereeing.","tokens_in":10820,"tokens_out":3005,"would_cite":true,"duration_ms":27124,"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":"KAN layers can replace linear layers in spoken-language understanding models, with a single KAN between two linear layers giving the best results.","keywords":["Kolmogorov-Arnold Networks","spoken language understanding","intent classification","transformer","wav2vec 2.0","XLS-R","B-spline","speech processing"],"falsifier":"Retrain the FFF, FFK, and FKF configurations on the same five datasets with at least ten random seeds and compare F1 distributions; if the FKF advantage over FFF overlaps zero within confidence intervals on the transformer backbones, the central claim of superior performance would collapse to 'comparable,' and on datasets where the point estimate favors FFF it would disappear.","tokens_in":9756,"feed_emoji":"🎙️","tokens_out":4316,"duration_ms":37673,"temperature":0.7,"pith_summary":"The paper asks whether Kolmogorov-Arnold Networks (KANs), which replace fixed activation functions with learnable univariate functions, can be substituted into the classifier head of spoken-language understanding (SLU) models. It reports that a single KAN layer placed between two linear layers (the FKF configuration) improves or matches accuracy and F1 on most of five intent-classification datasets, over both CNN and transformer backbones, with no meaningful training-time increase. The claim is architectural: one localized substitution in the classification head yields a small but consistent gain. A reader would care because SLU systems power voice assistants and smart devices, and any free accuracy from a drop-in layer swap is practically attractive.","feed_headline":"KAN layer between two linear layers wins in speech intent tests","feed_subtitle":"Swapping one classifier layer for a Kolmogorov-Arnold layer improves F1 on most of five spoken-language datasets.","key_machinery":"The key object is the KAN layer, a neural layer whose learnable univariate functions are implemented as a weighted combination of a basis function (Swish) and B-splines, per the efficient approximation in Eq. (2). The Kolmogorov-Arnold representation theorem motivates the layer: any multivariate continuous function can be written as sums of composed univariate functions, so KANs aim to learn these univariate pieces directly. The paper's central design choice is the FKF configuration, which places exactly one KAN layer between two linear feed-forward layers; this embedding lets linear layers perform coarse projection while the KAN captures localized nonlinear structure. The authors also use a recomputation trick that avoids expanding the input tensor to shape $(B, d_{out}, d_{in})$, turning the KAN computation into matrix multiplication and keeping memory and speed close to the linear baseline.","core_discovery":"The central discovery is that KAN layers are a viable drop-in replacement for linear layers in SLU classification heads, provided they are embedded rather than used alone. Across five datasets and two backbone families, the configuration with one KAN layer sandwiched between two feed-forward layers (FKF) achieves the best or near-best F1 macro scores, beating the all-linear baseline on FSC, Timers and Such, ITALIC, and SPEECH-MASSIVE, and matching it on SLURP. A KAN-only head performs worse than the linear baseline while adding parameters, showing that the benefit comes from combining KAN nonlinearity with linear projections. The authors also show that replacing the final-layer attention patterns changes which words drive predictions, making errors more human-plausible in their examples.","pith_inferences":["If the central claim holds, a natural next step is to test FKF on larger SLU benchmarks or on out-of-distribution speech to see whether the small gain persists when the classifier head has more capacity.","The observed shift in word-level attention could be studied as a regularizing side effect; one could test whether FKF improves robustness to background noise or mispronunciations.","The paper's comparisons are on intent classification only; the same dense-block substitution could be tried on slot filling or dialogue act classification, where the head structure is similar."],"forward_implications":["If the FKF configuration is adopted, SLU systems can gain small F1 improvements without increasing model size or training time.","The FKF pattern transfers from 2D-CNNs to wav2vec 2.0 and XLS-R transformers, suggesting the placement finding is architecture-agnostic.","B-spline basis functions are the best among the tested approximations for this task, so future KAN-SLU work should start with B-splines.","The attention analysis indicates that KAN-equipped heads attend to more semantically relevant words, which could make errors easier to interpret."],"supporting_citations":[{"why":"Defines KANs and the basis-plus-B-spline approximation that the paper builds on.","marker":"[1]"},{"why":"Provides the FLUENT SPEECH COMMANDS dataset used for 2D-CNN and transformer evaluation.","marker":"[19]"},{"why":"Provides the Timers and Such dataset used in ablation and attention analyses.","marker":"[20]"},{"why":"Supplies the wav2vec 2.0 backbone for the monolingual transformer experiments.","marker":"[26]"},{"why":"Supplies the XLS-R backbone for the multilingual transformer experiments.","marker":"[27]"},{"why":"Provides SLURP, the larger English dataset.","marker":"[28]"},{"why":"Provides ITALIC, the Italian dataset.","marker":"[29]"},{"why":"Provides SPEECH-MASSIVE for German and French.","marker":"[30]"},{"why":"Gives the word-level explanation technique used in the attention analysis.","marker":"[31]"}],"fun_headline_variants":["KAN layer between linear layers wins on speech intent tests","Sandwiched KAN layer beats linear baseline in most SLU settings","KAN as drop-in for linear layer improves F1 on 4 of 5 SLU sets","Best KAN setup: one KAN layer between two linear layers in speech","KAN layers replace linear heads effectively in spoken language understanding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported F1 differences between the FKF configuration and the all-linear baseline are stable across random initializations and not just noise; the paper does not state how many training runs each number averages over.","fun_headline_variants_meta":{"raw":{"variants":["KAN layer between linear layers wins on speech intent tests","Sandwiched KAN layer beats linear baseline in most SLU settings","KAN as drop-in for linear layer improves F1 on 4 of 5 SLU sets","Best KAN setup: one KAN layer between two linear layers in speech","KAN layers replace linear heads effectively in spoken language understanding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2855,"prompt_tokens":846,"completion_tokens":2009,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":1912}},"tokens_in":462,"tokens_out":2009,"duration_ms":28914,"temperature":1.0,"reasoning_tokens":1912,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:56:25.333387+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the FFF, FFK, and FKF configurations on the same five datasets with at least ten random seeds and compare F1 distributions; if the FKF advantage over FFF overlaps zero within confidence intervals on the transformer backbones, the central claim of superior performance would collapse to 'comparable,' and on datasets where the point estimate favors FFF it would disappear.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines KANs and the basis-plus-B-spline approximation that the paper builds on."},{"cited_title":"Tur and R","cited_arxiv_id":null,"evidence_quote":"Provides the FLUENT SPEECH COMMANDS dataset used for 2D-CNN and transformer evaluation."},{"cited_title":"A con- trastive learning approach to mitigate bias in speech models,","cited_arxiv_id":null,"evidence_quote":"Provides the Timers and Such dataset used in ablation and attention analyses."},{"cited_title":"Speech model pre-training for end-to-end spoken language understanding,","cited_arxiv_id":null,"evidence_quote":"Supplies the wav2vec 2.0 backbone for the monolingual transformer experiments."},{"cited_title":"Timers and such: A practical benchmark for spoken language un- derstanding with numbers,","cited_arxiv_id":null,"evidence_quote":"Supplies the XLS-R backbone for the multilingual transformer experiments."},{"cited_title":"Introduction to radial basis function networks,","cited_arxiv_id":null,"evidence_quote":"Provides SLURP, the larger English dataset."},{"cited_title":"Fasterkan,","cited_arxiv_id":null,"evidence_quote":"Provides SPEECH-MASSIVE for German and French."},{"cited_title":"Properties of Chebyshev polynomials","cited_arxiv_id":"2002.01342","evidence_quote":"Gives the word-level explanation technique used in the attention analysis."}],"review_version":1}