{"id":"f848e1b8-e964-45f7-9dcb-3219fbb803fa","arxiv_id":"2412.17778","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"On VoiceBank-DEMAND, replacing dense layers or ReLU activations with GR-KAN layers in MP-SENet and Demucs improved PESQ by up to 0.1 and cut parameters by up to 4x in one comparison.","lead":"GR-KAN, a variant of Kolmogorov-Arnold networks that uses shared rational functions, was plugged into MP-SENet and Demucs speech enhancement models. It modestly improved objective speech quality scores while shrinking model size, but the headline gains rest on a single dataset and partially incomplete reporting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Demucs parameter counts are identical after inserting parameterized GR-KAN activations, so the '4x fewer parameters' headline likely rests on omitted coefficients or a cross-depth comparison.","rationale":"The reader's weakest-assumption flags the absence of variance reporting for Demucs, and that is a real concern. I think the identical parameter counts are even more load-bearing because they can invalidate the quantitative headline even if the PESQ gains are perfectly reproducible. Replacing a zero-parameter ReLU with a learnable rational activation cannot leave total trainable parameters unchanged unless those activation parameters are not counted. The paper's abstract merges two separate comparisons: the 4x parameter reduction is depth-5 GR-KAN versus depth-6 baseline, while the up-to-0.1 PESQ improvement is depth-5 GR-KAN versus depth-5 baseline at equal parameter count. Neither comparison supports both claims together. This is a concrete, checkable issue in the reported numbers rather than a stylistic concern, so it should be resolved before the central claim is accepted as stated. I still recommend CONDITIONAL rather than REJECT because a parameter audit and a few seeded runs could straightforwardly confirm or refute the claim; the paper has honest implementation details and no internal inconsistency beyond this reporting gap.","tokens_in":8624,"tokens_out":5878,"duration_ms":60009,"concrete_test":"Audit all trainable parameters in the released implementation: sum p.numel() over every leaf parameter of the depth-5 baseline and GR-KAN Demucs, including every coefficient inside the GR-KAN rational functions. If the GR-KAN count is strictly greater than 18.868M, then Table 4's identical #P entries exclude the rational activation coefficients, and the '4x fewer parameters' claim must be recomputed from the audited counts. Print a per-module breakdown so the rational coefficients are visible and the exact parameter ratio can be verified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Table 4 reports exactly the same #P for baseline and GR-KAN at every Demucs depth (4.702, 18.868, 75.512M), yet Section 3.3 says zero-parameter ReLU activations were replaced by parameterized GR-KAN rational activations. If the rational-function coefficients are trainable, the GR-KAN model must have strictly more trainable parameters than the baseline, so the identical numbers strongly suggest those coefficients were excluded from #P. The headline 'up to 4x fewer parameters while improving PESQ by up to 0.1' is then unverified: the 4x ratio comes from comparing GR-KAN depth 5 (18.868M) with original depth 6 (75.512M), whereas the +0.1 PESQ gain comes from comparing GR-KAN depth 5 with original depth 5 at equal parameter count (2.896 -> 2.990). No single comparison in the paper demonstrates both claims simultaneously, and once any omitted activation parameters are counted, the parameter advantage may shrink or disappear. This is load-bearing because the central contribution is explicitly parameter efficiency plus quality gain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes adapting Group-Rational KAN (GR-KAN), a rational-function variant of Kolmogorov-Arnold Networks, to two speech enhancement architectures: MP-SENet in the time-frequency domain and Demucs in the time domain. The authors first compare KAN and GR-KAN against MLP variants on a synthetic signal modeling task, then replace dense layers in MP-SENet with GR-KAN layers and ReLU activations in Demucs with GR-KAN activations. On VoiceBank-DEMAND they report that GR-KAN improves PESQ by up to 0.1 while requiring up to 4x fewer parameters than a deeper baseline, and claim the first successful use of KAN-based methods for consistent improvement in both time-domain and state-of-the-art time-frequency-domain speech enhancement.","tokens_in":8849,"tokens_out":3685,"duration_ms":35586,"significance":"If the reported gains are reproducible and the parameter counts are correct, the paper offers a practical drop-in replacement for dense layers and activations in existing SE models, with modest objective quality improvements at lower or equal parameter cost. The paper has clear strengths: the MP-SENet experiments are averaged over three runs with standard deviations, the comparison includes both conventional and learnable activations, and the negative result for KAN on MP-SENet is a useful data point consistent with prior scalability concerns. However, the central parameter-efficiency claim currently rests on an apparent inconsistency in the Demucs parameter accounting, and the Demucs results lack variance information, so the headline improvement could be within run-to-run noise. The circularity burden is minimal, as GR-KAN is an externally proposed method and all comparisons are against external baselines.","major_comments":[{"comment":"Table 4 reports the same #P for the original Demucs and the GR-KAN adapted Demucs at every depth (4.702, 18.868, 75.512M), yet Section 3.3 states that the zero-parameter ReLU activations in the encoders and decoders were replaced by GR-KAN activations (GR from Eq. 4), and GR-KAN rational functions have trainable coefficients. If those coefficients are trainable, the GR-KAN model must have strictly more trainable parameters than the baseline. The identical numbers strongly suggest that the rational-function coefficients were excluded from #P. Please clarify exactly which parameters are included in the reported counts, and if the coefficients were omitted, recount and recompute all parameter-efficiency conclusions.","section":"Section 3.3 and Table 4"},{"comment":"The abstract's claim that 'GR-KAN requires up to 4x fewer parameters while improving PESQ by up to 0.1' is not supported by any single comparison in Table 4. The 4x ratio comes from comparing GR-KAN depth 5 (18.868M) with original depth 6 (75.512M), while the +0.1 PESQ improvement (2.896 to 2.990) comes from comparing GR-KAN depth 5 with original depth 5 at identical reported parameter counts. No row or pair of rows demonstrates both a 4x parameter reduction and a 0.1 PESQ gain simultaneously. The authors should either report a matched experiment that shows both outcomes in one comparison or rephrase the abstract to state the two findings separately and accurately.","section":"Abstract and Table 4"},{"comment":"The Demucs results in Tables 3 and 4 are reported as point estimates without standard deviations, confidence intervals, or the number of seeds, unlike Table 2 where MP-SENet results are averaged over three runs with standard deviations. Since the headline gains are approximately 0.1 PESQ (e.g., 2.896 to 2.990 at depth 5, and 2.977 to 3.018 at depth 6), it is currently impossible to judge whether these differences exceed run-to-run variation. Please report multi-run statistics for the Demucs experiments, or at minimum provide a variance estimate and a significance test, before claiming consistent improvement.","section":"Tables 3 and 4"}],"minor_comments":[{"comment":"There is a typo: 'V oicebank-DEMAND' should be 'VoiceBank-DEMAND'.","section":"Abstract"},{"comment":"The text uses 'scaler' where 'scalar' is intended; please correct the wording.","section":"Section 2.2, Eq. (3)"},{"comment":"The phrase 'To further access scalability' should read 'To further assess scalability'.","section":"Section 3.3"},{"comment":"The column headers 'KAN Enc' and 'KAN Dec' are misleading because the adapted model uses GR-KAN activations, not KAN activations; rename the headers to 'GR-KAN Enc' and 'GR-KAN Dec'.","section":"Table 3"},{"comment":"The Demucs adaptation uses only the GR rational-function part of Eq. (4) rather than the full GR-KAN layer; please specify the group size, rational-function degree, and any other GR-KAN hyperparameters used for the Demucs experiments, as these are not reported.","section":"Section 3.3"},{"comment":"The claim of 'first successful use of KAN-based methods' is stronger than the evidence supports, given that only two model architectures are tested and no comparison is made against other KAN variants or other recent activation-replacement methods on the same benchmark; consider softening the novelty claim.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The identical parameter counts in Table 4 are a serious red flag: replacing zero-parameter ReLU activations with trainable rational functions cannot leave the parameter count unchanged. I suspect the authors counted only the backbone architecture parameters and omitted the rational-function coefficients; this must be resolved before the paper can be considered further. The Demucs experiments also need variance information; without it, the central +0.1 PESQ claim is not verifiable. The MP-SENet portion of the paper is more solid and could form the basis for a revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing in this paper is the empirical message: plain KAN degrades MP-SENet, while GR-KAN gives small but consistent objective gains in both a TF-domain and a time-domain SE model on VoiceBank-DEMAND. That is a real, checkable result, and it extends the previous failed KAN attempt in a sensible way. The MP-SENet comparison is done carefully: 3 runs, standard deviations, and multiple activation baselines including learnable ones. The small-scale signal fitting experiment is a nice sanity check.\n\nThe soft spots are real but fixable. First, the Demucs parameter counts: Table 4 reports exactly the same #P for baseline and GR-KAN at every depth, yet Section 3.3 says the zero-parameter ReLUs were replaced with parameterized rational activations. Either those coefficients are not counted, which makes the 'fewer parameters' claim inflated, or there is a typo. The authors should show the breakdown. Second, the abstract's 'up to 4x fewer parameters while improving PESQ by up to 0.1' combines two different comparisons: the 4x comes from comparing GR-KAN depth 5 with original depth 6, while the +0.1 comes from comparing GR-KAN depth 5 with original depth 5. It is not false—there is a single comparison (GR-KAN depth 5 vs original depth 6) that shows both a 4x parameter reduction and a small PESQ gain—but the wording is misleading. Third, the Demucs results in Tables 3 and 4 report no variance or seed count, so the +0.1 gain could be noise. Table 2 shows that MP-SENet gains are small but outside one standard deviation; the Demucs numbers need the same treatment. Minor: the KAN models in MP-SENet were trained with batch size 2 versus 4 for the others, which confounds the KAN comparison, though the GR-KAN versus dense comparison is unaffected.\n\nThe citation pattern looks honest: they cite the KAN and GR-KAN sources and the prior failed MetricGAN+ application. The core architecture is prior work, so novelty is moderate—this is a domain application with new measurements, not a new method.\n\nWho it's for: people in speech enhancement who are considering KAN-style activations. It deserves a serious referee; the flaws are addressable and the empirical result is potentially useful.\n\nRecommendation: send to peer review, with a request that the authors clarify the parameter accounting and add variance reporting for Demucs.","headline":"Useful empirical application of GR-KAN to speech enhancement, but the Demucs parameter accounting is inconsistent and the headline mixes two comparisons.","tokens_in":9405,"tokens_out":3794,"would_cite":false,"duration_ms":32151,"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":"GR-KAN, a group-rational variant of Kolmogorov-Arnold Networks, improves speech enhancement quality in both time- and time-frequency-domain models while reducing parameters by up to 4x.","keywords":["speech enhancement","Kolmogorov-Arnold Networks","GR-KAN","rational activation functions","time-frequency domain speech enhancement","time-domain speech enhancement","parameter efficiency","PESQ"],"falsifier":"Train the original and GR-KAN-adapted Demucs at depth 5 on VoiceBank-DEMAND across ten random seeds; if the mean PESQ difference between 2.990 and 2.896 falls below the standard deviation of the baseline, the claimed consistent improvement is not established.","tokens_in":8441,"feed_emoji":"🎧","tokens_out":5789,"duration_ms":49583,"temperature":0.7,"pith_summary":"This paper argues that a scalable variant of Kolmogorov-Arnold Networks, called Group-Rational KAN (GR-KAN), can be dropped into existing speech enhancement models to improve objective speech quality while using far fewer parameters. The authors replace dense layers in the time-frequency-domain MP-SENet and ReLU activations in the time-domain Demucs with GR-KAN layers, and report consistent gains on the VoiceBank-DEMAND benchmark: up to 0.1 higher PESQ and up to four times fewer parameters. They also show that the original KAN, with spline activations on every edge, fails to scale to this task, adding about 50% parameters without improving MP-SENet. If these results hold, GR-KAN offers a plug-in path to better speech enhancement without redesigning the network.","feed_headline":"Group-rational KAN nets lift speech quality at 4x fewer parameters","feed_subtitle":"A drop-in KAN variant improves both time- and frequency-domain models on the VoiceBank-DEMAND benchmark.","key_machinery":"The load-bearing object is the GR-KAN layer, implemented as $\\mathrm{LIN}(\\mathrm{GR}(x))$: a group-wise rational activation $\\mathrm{GR}$ followed by a linear matrix $\\mathrm{LIN}$. Unlike a KAN layer, which places an independent spline on each edge, GR-KAN partitions input channels into $k$ groups and shares one rational function per group, which keeps parameters low, and it uses variance-preserving weight initialization for stable training. In MP-SENet the layer is inserted after the Bi-GRU inside each GRU-Transformer block; in Demucs its rational function replaces the ReLU nonlinearity in 1D CNN encoder/decoder blocks. This mechanism supplies the extra expressiveness for modeling harmonic and phase structure while avoiding the parameter blow-up that makes plain KAN fail.","core_discovery":"The central claim is that GR-KAN layers—learnable rational functions shared across channel groups, followed by a linear projection—are expressive enough to capture the smooth, multiscale structure of speech, while remaining trainable at scale. On MP-SENet, replacing the dense layers inside the GRU-Transformer blocks with one GR-KAN layer raises PESQ from 3.561 (GELU baseline) to 3.588 at the same 2.26M parameters, and outperforms doubled dense-layer baselines. On causal Demucs, replacing ReLU with GR-KAN in the encoder/decoder lifts PESQ from 2.896 to 2.990 at depth 5, matching the depth-6 original, which has over four times the parameters. The paper presents this as the first consistent improvement from KAN-based methods in both time- and state-of-the-art time-frequency-domain speech enhancement.","pith_inferences":["Our inference: the parameter savings suggest GR-KAN could be tested in larger speech enhancement backbones or in vocoders, where replacing expensive activations would be cheap to try.","Our inference: the group-sharing mechanism, rather than rational functions alone, may be the key to the result; an ablation that uses a rational-family activation in the same slots would isolate that factor.","Our inference: the method should transfer to non-speech audio restoration, such as music denoising, if the expressiveness claim about multiscale structure is general."],"forward_implications":["The same training recipes and loss functions as the original MP-SENet and Demucs can be kept; GR-KAN is a drop-in swap for dense layers or activations.","At depth 5, the GR-KAN-adapted Demucs matches the PESQ of the original depth-6 model while using about a quarter of the parameters, so parameter budgets can be redirected to depth or other capacity.","GR-KAN beats doubled dense-layer baselines in MP-SENet, suggesting that more parameters with conventional layers are not an equivalent substitute for the rational activation structure.","Plain KAN does not help time-frequency-domain speech enhancement at scale, so future KAN-style speech enhancement work should use GR-KAN's group-rational design rather than spline KAN.","The paper concludes that future speech enhancement and speech generation systems may benefit from adopting GR-KAN."],"supporting_citations":[{"why":"Defines KAN layers with per-edge spline activations, the expressiveness baseline that GR-KAN modifies.","marker":"[27]"},{"why":"Introduces Group-Rational KAN, including rational functions, group sharing, and variance-preserving initialization that the paper adopts.","marker":"[28]"},{"why":"Demucs, the time-domain model whose ReLU activations are replaced by GR-KAN.","marker":"[9]"},{"why":"MP-SENet, the time-frequency-domain state-of-the-art model whose dense layers are replaced.","marker":"[20]"},{"why":"Prior KAN-for-speech-enhancement attempt in MetricGAN+ that degraded performance, motivating GR-KAN.","marker":"[30]"},{"why":"VoiceBank-DEMAND, the benchmark dataset for all speech enhancement experiments.","marker":"[33]"},{"why":"PESQ, the primary perceptual quality metric whose gains anchor the main claims.","marker":"[37]"},{"why":"Kolmogorov-Arnold theorem, the theoretical basis for representing functions as sums of univariate functions.","marker":"[32]"}],"fun_headline_variants":["GR-KAN beats dense layers at quarter the size","KAN variant boosts speech quality with fewer params","Group-rational KAN sharpens speech enhancement","GR-KAN: smaller nets, clearer speech","KAN upgrade yields clearer speech with 4x fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported Demucs gains, especially the +0.1 PESQ from depth 5, are larger than run-to-run variation; the paper gives no standard deviations or seed count for those tables.","fun_headline_variants_meta":{"raw":{"variants":["GR-KAN beats dense layers at quarter the size","KAN variant boosts speech quality with fewer params","Group-rational KAN sharpens speech enhancement","GR-KAN: smaller nets, clearer speech","KAN upgrade yields clearer speech with 4x fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000464,"raw_usage":{"total_tokens":2317,"prompt_tokens":940,"completion_tokens":1377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1302}},"tokens_in":556,"tokens_out":1377,"duration_ms":8916,"temperature":1.0,"reasoning_tokens":1302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:14.832070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the original and GR-KAN-adapted Demucs at depth 5 on VoiceBank-DEMAND across ten random seeds; if the mean PESQ difference between 2.990 and 2.896 falls below the standard deviation of the baseline, the claimed consistent improvement is not established.","supporting_citations":[{"cited_title":"A regression approach to speech enhancement based on deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Demucs, the time-domain model whose ReLU activations are replaced by GR-KAN."},{"cited_title":"On the representation of continuous func- tions of many variables by superposition of continuous functions of one variable and addition,","cited_arxiv_id":null,"evidence_quote":"PESQ, the primary perceptual quality metric whose gains anchor the main claims."}],"review_version":1}