{"id":"28e86411-a46e-4875-8d19-de0c7e5ef838","arxiv_id":"2412.19713","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A progressive KAN-stacking method is claimed to improve liver tumor segmentation accuracy and efficiency, yet its own reported results are internally inconsistent.","lead":"This paper proposes proKAN, a network that adds Kolmogorov-Arnold blocks during training only when validation performance stalls, aiming to limit overfitting and compute cost. It reports higher Dice and accuracy than plain MLPs and fixed KANs on liver tumor segmentation, but the supporting tables contain contradictions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Dice scores are mutually inconsistent (92.3% in Tables III/IV vs 0.58 in Table VI), so the central claim of state-of-the-art segmentation performance has no coherent experimental support.","rationale":"The reader's weakest_assumption focuses on the progressive stacking rule as a plausible but unsupported mechanism for mitigating overfitting. That is a reasonable concern, but I identify a more fundamental and more easily settled problem: the reported quantitative evidence for the central claim is internally contradictory. Tables III and VI give Dice scores for proKAN that differ by a factor of about 159 (92.3% vs 0.58), and Table IV contains values identical to PVTFormer, suggesting the numerical support has not been carefully validated. Because the paper's headline contribution is empirical ('outperforming standard MLPs and fixed KAN architectures'), inconsistent numbers undermine the central claim regardless of whether the progressive stacking idea is theoretically sound. The concrete test I propose would resolve the contradiction with minimal effort: recompute Dice from per-fold predictions. No change to the reader's REJECT verdict is warranted; if anything, the inconsistency strengthens the rejection. My agreement is 'partial' because the reader did also mention the Dice discrepancy in their rationale, but they did not make it the load-bearing weakest assumption. I do not raise ad hominem concerns; the issue is the evidence, not the authors' intent.","tokens_in":8854,"tokens_out":2381,"duration_ms":24908,"concrete_test":"Request the per-fold Dice scores, predicted segmentation masks, and ground-truth labels from the authors (or independently re-run the 10-fold cross-validation on LiTS17 using the described proKAN architecture). Recompute DSC using the definition in Section IV-B and compare each fold's value to Tables III, IV, and VI. If the recomputed Dice for proKAN is approximately 0.58, then Table III's 92.3% is erroneous and the outperformance claim fails. If the recomputed Dice is approximately 0.923, then Table VI's 0.58 is erroneous and the overfitting-mitigation and segmentation-quality analysis in Section IV-D is invalid. In either case, the paper must be revised to present one consistent, reproducible metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that proKAN 'achieves state-of-the-art performance in liver segmentation tasks' rests on the quantitative results in Tables III, IV, and VI. These tables are internally inconsistent. Table III reports proKAN Dice = 92.3%, MLP = 85.2%, and KAN = 89.5%. Table IV repeats proKAN Dice = 92.3% but also lists proKAN mIoU = 78.46 and HD = 3.50, which are identical to the values for PVTFormer in the same table. Table VI reports the average Dice score over the 10-fold cross-validation as 0.58 for proKAN, 0.57 for KAN, and 0.55 for MLP. Since the Dice Similarity Coefficient is defined in Section IV-B as DSC = 2|X ∩ Y|/(|X| + |Y|), a bounded measure in [0,1], the value 0.58 and the value 92.3% cannot both describe the same model on the same benchmark. If Table VI is correct, proKAN does not outperform the baselines in Dice; if Table III is correct, Table VI is wrong. Either way, the paper does not present a consistent set of experimental results, and the claimed empirical superiority of proKAN is unsupported. This is not a matter of interpretation or a weak ablation; it is a mechanical contradiction within the paper's own reported evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces proKAN, a progressive stacking method for Kolmogorov-Arnold Networks intended for liver tumor segmentation. The method starts with a small KAN-enhanced MLP and adds KAN blocks during training when validation loss plateaus while training loss continues to decrease, with associated dynamic adjustments to grid size, spline degree, learning rate, and regularization. The authors report experiments on the LiTS17 dataset comparing proKAN against MLPs and fixed KANs, and claim state-of-the-art accuracy, Dice score, and training time, as well as improved interpretability and overfitting mitigation.","tokens_in":9218,"tokens_out":2052,"duration_ms":22305,"significance":"If the reported results were reliable, the idea of dynamically growing a KAN's capacity based on validation behavior would be a useful contribution to efficient medical image segmentation. However, the central quantitative claim is undermined by internal contradictions in the paper's own tables, as detailed below. The paper also provides no learning curves, no ablation of the progressive-stacking rule, no code or reproducibility artifacts, and no comparison against standard early stopping, so the overfitting-mitigation claim is not currently supported. The method itself is not inherently implausible, but the experimental evidence as presented does not justify acceptance.","major_comments":[{"comment":"The reported Dice scores for proKAN are mutually inconsistent. Table III reports proKAN Dice Score = 92.3% (with MLP 85.2% and KAN 89.5%), while Table VI reports the average Dice score over 10-fold cross-validation as 0.58 for proKAN, 0.57 for KAN, and 0.55 for MLP. Since DSC is defined in Section IV-B as a measure in [0,1], either the Table III values are percentages (0.923) or the Table VI values are decimals (0.58), and the two cannot describe the same model on the same benchmark. This is a mechanical contradiction, not a matter of interpretation; the paper does not provide a coherent quantitative basis for the claim that proKAN outperforms both baselines in Dice score.","section":"Tables III and VI"},{"comment":"In Table IV, proKAN reports mIoU = 78.46 and HD = 3.50, which are exactly identical to the values reported for PVTFormer in the same table, even though the Dice values differ (92.3% vs. 86.78%). Exact equality of two independent evaluation metrics across two different models is not impossible, but it is highly implausible without comment. The paper gives no explanation for this coincidence, and the row therefore raises a serious concern that these metrics were not independently computed for proKAN. This directly affects the state-of-the-art comparison in the abstract and Section IV-C.","section":"Table IV"},{"comment":"The central methodological premise is that adding KAN blocks when validation loss plateaus (while training loss continues to decrease) mitigates overfitting. This premise is not supported by any evidence in the paper. No learning curves are shown, no ablation is performed, no comparison with standard early stopping is given, and the 10-fold cross-validation tables only report final accuracy and Dice, not the evolution of the training/validation gap. The paper also uses the validation set both to decide when to add blocks and to report the headline validation accuracy, creating a selection-on-validation effect that is not discussed. These issues are load-bearing because the overfitting-mitigation claim is one of the two main advertised contributions.","section":"Section III-B and Section IV-D"}],"minor_comments":[{"comment":"The abstract and introduction claim that proKAN is designed for real-time clinical deployment on FPGAs and ASICs, but Section IV-F only reports inference speed and GPU memory usage; no FPGA or ASIC experiments or analyses are presented.","section":"Abstract and Section I"},{"comment":"The term 'KAN binding' appears in the introduction and related work, but it is never formally defined or used in the methodology; Section III only describes KAN blocks and progressive stacking.","section":"Section II and Section III"},{"comment":"The caption of Table IV cites reference [2] (Hame and Pollari) as the benchmark dataset, but the dataset is LiTS17, and [20] is the actual LiTS benchmark reference; the citation appears to be incorrect.","section":"Table IV caption"},{"comment":"The text says 'Table III summarizes the performance' and claims proKAN mitigates overfitting compared to MLPs and KANs, but Table III only reports validation accuracy, training time, and Dice score; it contains no measure of overfitting such as train-validation gap.","section":"Section IV-C"},{"comment":"The dynamic hyperparameter adjustments (grid size, spline degree, learning rate, regularization) are stated with free parameters ΔG, Δk, α, and Δλ, but no values for these parameters are given anywhere in the paper, making the experiments irreproducible.","section":"Section III-C and Table II"}],"recommendation":"reject","confidential_remarks":"The internal contradictions in Tables III, IV, and VI are severe enough that I cannot recommend revision within the normal scope. In particular, the exact duplication of PVTFormer's mIoU and HD values in Table IV warrants editorial scrutiny of the underlying experimental records."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the progressive-stacking idea is worth a sentence, but the paper's own tables contradict each other, and the quantitative claims collapse. Table III gives proKAN Dice 92.3%; Table VI reports average Dice 0.58 for the same model. Table IV gives proKAN mIoU 78.46 and HD 3.50, identical to PVTFormer in the same table, which suggests the numbers were copied rather than computed. That is not a quibble; it is the whole empirical basis.\n\nWhat is new: combining progressive depth growth with KAN spline activations for liver segmentation is a reasonable idea. The related work covers CNNs and hybrid 2D/3D networks, and dynamic-depth heuristics exist, so the contribution is incremental but not zero. The writing is clear enough, and the method section states the plateau rule and hyperparameter adjustments directly.\n\nWhat is soft: the experimental inconsistency alone is enough to reject. There are also no error bars, no learning curves, no ablation of the progressive rule against fixed early stopping, and no code. The progressive stacking rule itself is questionable: when validation loss plateaus while training loss falls, adding more capacity normally worsens overfitting rather than curing it. The paper does not show that it behaves otherwise. The original KAN paper is not cited, which is a serious omission for a method built on it. The hyperparameters are numerous (grid increment, spline degree increment, learning-rate decay, regularization increment, plateau threshold and window) and there is no sensitivity analysis.\n\nCredit where due: the prose is understandable and the limitation of fixed networks is a real problem. But the paper does not provide a coherent set of numbers. The stress-test note is right: 0.58 and 92.3% cannot describe the same Dice coefficient on the same benchmark.\n\nWho this is for: readers curious about KAN variants in segmentation might skim the method section, but nobody should rely on the results. The paper does not deserve a serious referee until the tables are reconciled, the experiments rerun, and code/data made available. My recommendation: desk reject with an invitation to resubmit once the evidence is fixed.","headline":"ProKAN's progressive-depth idea is plausible, but the paper's own tables contradict each other, so the empirical claims collapse.","tokens_in":9701,"tokens_out":1989,"would_cite":false,"duration_ms":19118,"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":"ProKAN claims a Kolmogorov-Arnold network that adds blocks only when validation loss plateaus beats fixed MLP and KAN baselines on liver tumor segmentation, reaching 96.7% validation accuracy and 92.3% Dice.","keywords":["Kolmogorov-Arnold Networks","progressive stacking","liver tumor segmentation","B-spline activations","overfitting mitigation","LiTS17","medical image segmentation","computational efficiency"],"falsifier":"Run a controlled comparison on LiTS17 where a fixed-depth KAN is trained with standard early stopping and the same total compute as proKAN; if the fixed KAN matches or beats proKAN on validation accuracy, Dice, and wall-clock time, the progressive rule is not responsible for the reported gains. Independently, recompute Dice from proKAN's predicted masks against ground truth, since the value should be near 0.92 if Tables III and IV are right and near 0.58 if Table VI is right.","tokens_in":8686,"feed_emoji":"🩻","tokens_out":6940,"duration_ms":66647,"temperature":0.7,"pith_summary":"This paper proposes proKAN, a training scheme that grows a Kolmogorov-Arnold Network (KAN) in depth during training instead of fixing the architecture in advance. The growth rule is: when validation loss stops improving while training loss keeps falling, add another KAN block; each addition also enlarges the spline grid, raises the spline degree, lowers the learning rate, and increases regularization. On the LiTS17 liver tumor segmentation benchmark, the authors report 96.7% validation accuracy, a 92.3% Dice score, and 5.5 hours of training, compared with 92.3%, 85.2%, and 4.0 hours for an MLP and 95.1%, 89.5%, and 8.0 hours for a fixed KAN. The paper interprets the smaller train-validation gap across 10-fold cross-validation as evidence that the progressive rule mitigates overfitting. If the numbers hold, proKAN would offer a practical self-sizing alternative to fixed-depth networks for volumetric medical segmentation.","feed_headline":"Self-growing KAN beats fixed networks on liver segmentation","feed_subtitle":"ProKAN reports 96.7% accuracy and 92.3% Dice on LiTS17 in 5.5 hours of training.","key_machinery":"The load-bearing mechanism is the progressive stacking rule for KAN blocks. A KAN block is a layer in which each connection is a learnable univariate function parameterised by a B-spline, so the network has no fixed activation shapes; the Kolmogorov-Arnold theorem licenses the decomposition of a multivariate function into sums of such univariate functions. proKAN starts with a base MLP augmented by one KAN block and inserts further blocks when the average validation loss over a plateau window is flat while training loss is still falling, or when validation accuracy has begun to decline. Each insertion adjusts the architecture's hyperparameters, with spline grid granularity and degree increasing, learning rate decaying, and regularization strengthening, so that added capacity is paired with stabilisation. The rule is what is supposed to convert KAN flexibility into efficient, overfitting-resistant segmentation.","core_discovery":"The central claim is that progressive stacking of KAN blocks, starting from a minimal network and adding capacity only when validation plateaus while training loss continues to decrease, produces higher liver segmentation accuracy, better Dice overlap, and shorter training time than both MLPs and fixed KANs. The authors frame this as a general overfitting-mitigation strategy: the network stops growing once the validation signal says further capacity would be wasted, so it avoids both underfitting and unnecessary computation. They attribute the flexibility of the building block to the Kolmogorov-Arnold representation theorem, with each connection carrying a learnable univariate B-spline activation rather than a fixed nonlinearity. In their experiments on LiTS17 CT scans, proKAN reaches 96.7% validation accuracy and a 92.3% Dice score; the authors also report it bests all seven previously published segmentation baselines listed on the same benchmark, with 15.3 ms per-image inference and 2.8 GB GPU memory usage.","pith_inferences":["The growth rule is the part most worth testing in isolation: a fixed KAN trained with standard early stopping is the natural control, since adding capacity when validation plateaus runs against the usual expectation that such a moment signals too much capacity, not too little.","The paper's two reported Dice values for proKAN, 92.3% in Tables III and IV versus 0.58 in Table VI, cannot both be right; reconciling them would clarify which claim the method actually supports.","If progressive stacking is validated, the same plateau-detection criterion could be applied to other architecture families, such as deciding when to unroll more transformer layers or add residual blocks, making it a generic capacity-scheduling rule rather than a KAN-specific trick."],"forward_implications":["If the progressive rule works, architecture depth becomes a training-time decision rather than a pre-committed hyperparameter, so the same code can serve small and large datasets without manual depth search.","The reported 5.5-hour training time against 8.0 hours for a fixed KAN suggests that a self-sizing KAN can recover most of the accuracy gain of KANs while avoiding their full computational cost.","The dynamic hyperparameter schedule, with grid size and spline degree increasing and learning rate decreasing as blocks are added, gives a concrete recipe for scaling KAN capacity that could transfer to other 3D segmentation tasks.","If the Dice and accuracy numbers reproduce, proKAN would be competitive with published LiTS liver segmentation methods while offering the interpretability of spline coefficients."],"supporting_citations":[{"why":"Supplies the LiTS17 contrast-enhanced CT dataset and ground-truth segmentations used for every reported experiment.","marker":"[20]"},{"why":"Provides the DoubleU-Net baseline that proKAN claims to exceed on Dice over the LiTS dataset.","marker":"[21]"},{"why":"Provides the UNeXt baseline on LiTS that proKAN claims to outperform in Dice and mIoU.","marker":"[24]"},{"why":"Provides the PVTFormer LiTS baseline against which proKAN's reported Dice is compared.","marker":"[27]"}],"fun_headline_variants":["KANs that self-grow only when needed win on liver scans","Adaptive KAN stops at overfitting, beats fixed nets on liver","Progressive KAN shrinks training time, lifts liver Dice score","Self-stopping KAN outperforms MLPs in 3D liver segmentation","Dynamic KAN builds itself, then halts to beat static rivals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim collapses if adding KAN blocks when validation loss plateaus while training loss falls does not actually reduce overfitting, and the paper provides no learning curves, ablations, or an early-stopping control to demonstrate that the rule helps rather than hurts.","fun_headline_variants_meta":{"raw":{"variants":["KANs that self-grow only when needed win on liver scans","Adaptive KAN stops at overfitting, beats fixed nets on liver","Progressive KAN shrinks training time, lifts liver Dice score","Self-stopping KAN outperforms MLPs in 3D liver segmentation","Dynamic KAN builds itself, then halts to beat static rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1677,"prompt_tokens":1023,"completion_tokens":654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":639,"tokens_out":654,"duration_ms":7689,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:56:10.338533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison on LiTS17 where a fixed-depth KAN is trained with standard early stopping and the same total compute as proKAN; if the fixed KAN matches or beats proKAN on validation accuracy, Dice, and wall-clock time, the progressive rule is not responsible for the reported gains. Independently, recompute Dice from proKAN's predicted masks against ground truth, since the value should be near 0.92 if Tables III and IV are right and near 0.58 if Table VI is right.","supporting_citations":[{"cited_title":"The liver tumor segmentation benchmark (lits),","cited_arxiv_id":null,"evidence_quote":"Supplies the LiTS17 contrast-enhanced CT dataset and ground-truth segmentations used for every reported experiment."},{"cited_title":"Doubleu-net: A deep convolutional neural network for medical im- age segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the DoubleU-Net baseline that proKAN claims to exceed on Dice over the LiTS dataset."},{"cited_title":"Unext: Mlp-based rapid medical image segmentation network,","cited_arxiv_id":null,"evidence_quote":"Provides the UNeXt baseline on LiTS that proKAN claims to outperform in Dice and mIoU."},{"cited_title":"CT Liver Segmentation via PVT-based Encoding and Refined Decoding","cited_arxiv_id":"2401.09630","evidence_quote":"Provides the PVTFormer LiTS baseline against which proKAN's reported Dice is compared."}],"review_version":1}