{"id":"c636e506-9ce6-4857-990a-b7b5aa80d05d","arxiv_id":"2608.12194","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HYDRA replaces dense KAN edge functions with a low-rank prototype block in a radius-constrained Poincaré ball, matching or beating KAN and MLP on eight tabular benchmarks with fewer parameters.","lead":"HYDRA is a new neural network architecture that combines Kolmogorov-Arnold Networks with hyperbolic geometry to learn compact, interpretable models. A smart generalist might read it to see whether hyperbolic latent spaces can reduce parameters in functional learning without sacrificing accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline comparison is not auditable: Table 1 parameter counts and metrics cannot be reproduced from Appendix D's stated widths/ranks, and all runs use one seed/split.","rationale":"The reader's weakest-assumption candidate was the low effective dimension of the tangent-space update, and the reader's rationale also noted inconsistent numbers and single-seed results. I agree that the low-rank assumption is under-supported, but I see a more load-bearing problem: the central empirical claim cannot be audited from the paper as written. Table 1's parameter counts are not derivable from the architecture hyperparameters in Appendix D with any stated layer-count or parameter formula, and the primary metrics shift across Table 1, Tables 2/3, and Table 5 for the same datasets and seed. These are internal inconsistencies, not merely disagreements with prior work. The universal approximation argument in Appendix A provides genuine theoretical support for the architecture, and the proposed radius-control mechanism is plausible, so the contribution may have value; however, the headline claim that HYDRA is both strongest and parameter-efficient is currently unverifiable. I would not reject the submission outright, but I would move it to unverdictable until the tables are reconciled and a variance-aware evaluation is supplied.","tokens_in":15456,"tokens_out":8892,"duration_ms":83921,"concrete_test":"Recompute every Table 1 parameter count from explicit layer counts and the formulas in Eqs. (8)-(9) using Appendix D hyperparameters; then rerun all eight datasets with 10 independent seeds/splits under the same selection rule and report mean±std and per-seed rankings. If the Table 1 counts do not match the formulas, or if HYDRA is not statistically better than the closest baseline on most datasets, the headline claim fails.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim depends entirely on Table 1, but Table 1 cannot be checked against the reported configuration. Appendix D gives HYDRA CCPP w=43, K=6, r=6; the low-rank block alone is 2dr + r^2(K+1) = 2*43*6 + 36*7 = 768 parameters before any embedding/readout or depth multiplier, while Table 1 reports 4.8k trainable parameters; depth/layer count is never specified. For KAN CCPP, Appendix D gives w=51, K=12; a d=51 hidden KAN block is 51^2*(12+1) = 33.8k parameters, far above the reported 7.0k, and even an input layer 4->51 is 2.65k. Since no formula or layer count is supplied, the parameter-efficiency comparison—the second half of the central claim—is not reproducible. The metrics are also internally inconsistent across tables for the same dataset/seed: Phoneme accuracy is 0.885 in Table 1 but 0.856/0.845 in Tables 3/5; Energy RMSE is 0.706 in Table 1 but 1.193/1.384 in Tables 3/5. The paper reports a single fixed split and seed 42 with no variance estimate, so the 'strongest on all eight' claim is compatible with selection over many configurations. The low-rank assumption noted by the reader is real, but the more immediate blocker is that the evidence table itself cannot be reconstructed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HYDRA is a proposed KAN variant that stores hidden states on the Poincaré ball, performs KAN-style spline updates in the tangent space through a low-rank prototype bottleneck, and adds a radius-control penalty and projection. The paper claims that across eight tabular benchmarks HYDRA achieves the strongest or tied-strongest primary metric on every dataset while using fewer trainable parameters than Euclidean KAN and MLP, and that ablations support the low-rank and radius-control design choices. Appendices provide a universal approximation argument, hyperparameter tables, additional metrics, and a stated limitations section.","tokens_in":15827,"tokens_out":8293,"duration_ms":78382,"significance":"If the empirical claims were fully supported, HYDRA would be a meaningful parameter-efficiency improvement for KAN-style architectures, and the radial-coordinate diagnostic would be a useful interpretability addition. The low-rank parameter count in Eq. (9)-(10) is a clean and correct design contribution, and the radius-control mechanism is a sensible response to known Poincaré-boundary instability. The universal approximation attempt is non-circular in intent because it is checked against the Euclidean KAN realization rather than against HYDRA's fitted values. However, the central empirical claim is currently not auditable: the reported parameter counts do not follow from the stated configurations, identical primary metrics differ across tables, and all results rest on a single split and seed. These issues, together with a gap in the scaling argument of Appendix A, mean that the paper's headline conclusions are not yet established.","major_comments":[{"comment":"The parameter-efficiency comparison cannot be reconstructed from the supplied configurations. For HYDRA on CCPP, Table 6 gives w=43, K=6, r=6, so a single low-rank block already contains 2dr + r^2(K+1) = 2*43*6 + 36*7 = 768 parameters; with a 4-dimensional linear embedding (172 parameters) and a linear readout (43 parameters), the reported 4.8k total is consistent only with an unreported number of blocks or additional structure. For KAN on CCPP, Table 7 gives w=51, K=12; a dense 51-to-51 KAN block has 51^2*(12+1) = 33,813 parameters, and even a single 4-to-51-to-1 KAN has 2,652 + 663 = 3,315 parameters, neither of which matches the reported 7.0k. Since no depth or layer count is given for any model, the claimed 34.9% parameter reduction relative to Euclidean KAN is not verifiable.","section":"Table 1 vs. Appendix D"},{"comment":"The same primary metric for the same model family takes different values in different tables. Energy Heating RMSE is 0.706 in Table 1 but 1.193 in Table 3 and 1.193/1.384 in Table 5; Phoneme accuracy is 0.885 in Table 1 but 0.856/0.845 in Tables 3 and 5. Moreover, the selected low-rank HYDRA results in Table 2 do not match Table 1 for several datasets, e.g., CCPP 3.919 vs. 3.604, Ionosphere 0.943 vs. 0.971, and QSAR 0.882 vs. 0.900. These are not auxiliary metrics; they are the same primary metrics used for the headline claim, so the reader cannot tell which configuration produced the reported performance.","section":"Tables 1, 2, 3, and 5"},{"comment":"All runs use a single fixed split and seed 42 with no standard deviations or repeated trials, yet the paper states that HYDRA 'achieves the strongest or tied-strongest primary metric on all eight datasets.' With a single split, and with rank and radius selected from multiple configurations as the paper itself acknowledges in Appendix E, the reported ranking is compatible with selection over many candidates. The claim requires either repeated runs with variance estimates or a clearly prespecified evaluation protocol, especially for datasets where the margins between models are small.","section":"Experimental Setup and Appendix E"},{"comment":"The universal approximation proof contains a scaling step that is not justified for spline KANs. The proof scales the Euclidean realization to z_l^(α) = α z_l, then defines f_HYDRA = α^{-1} Q H^(α) and g_ϕ = α^{-1} Q G^(α). This assumes that rescaling the hidden coordinates rescales the Euclidean network's output by α, i.e., G^(α) = α g_ϕ, but KAN spline blocks are not homogeneous: φ(α z_j) is not generally α φ(z_j). The displayed O(R_α^3) bounds control only the exp/log discrepancy, not the discrepancy between the rescaled Euclidean network and the original g_ϕ. A repair would need to construct a Euclidean network with bounded hidden states directly and absorb all output scaling into the final readout.","section":"Appendix A, Eqs. (A6)-(A11)"}],"minor_comments":[{"comment":"The notation for layer-dependent quantities is introduced loosely: the number of blocks L, the residual scale α_l, the radius budget r_l, and the soft threshold r_allow,l are used before their schedules are specified. Please state the layer-count and radius-budget definitions explicitly.","section":"Method, Eq. (1)"},{"comment":"The caption says that the smallest and second-smallest parameter counts are bold and underlined, but no bold or underline appears inside the parentheses in the table as typeset. The formatting should be corrected or the caption amended.","section":"Table 1 caption"},{"comment":"The sentence following Eq. (5) contains a grammatical error: 'that leading to every input-output coordinate pair owns a separate spline' should be rephrased, e.g., 'so that every input-output coordinate pair owns a separate spline.'","section":"Eq. (5)"},{"comment":"The text refers to panel (b) in Figure 2 and to specific visual effects in Figure 3, but the figures appear only as placeholders in the manuscript. Please ensure the final figures are legible and that panels are labeled consistently with the text.","section":"Figures 2 and 3"}],"recommendation":"major_revision","confidential_remarks":"The idea is plausible and the low-rank bottleneck is a natural contribution, but the paper is not in a publishable state until the parameter counts are made reproducible, the metric inconsistencies across tables are resolved, and repeated-seed results are provided. I found no indication of misconduct; the issues are internal consistency and auditability. The Appendix A proof also needs a genuine repair rather than a cosmetic edit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHYDRA is a KAN variant that combines a low-rank prototype bottleneck with hyperbolic (Poincaré) latent states and radius control. The core idea is sensible and new: the low-rank spline update in tangent space replaces the O(d^2 K) hidden-to-hidden spline cost with O(dr + r^2 K), and the radius penalty is a reasonable counter to boundary saturation. The universal approximation proof in Appendix A is standard but correct: any Euclidean KAN can be embedded in a full-rank HYDRA by taking r >= d and scaling inputs toward the origin. Note that this density result does not cover the low-rank regime (r < d) that the paper actually uses; the paper admits this, calling low-rank HYDRA a subfamily of the universal class. That is honest, but it means the efficient version has no universality guarantee.\n\nWhat the paper does well: the ablations are well motivated. The low-rank ablation compares against full-rank HYDRA, and the radius-control ablation shows that constrained models have smaller mean radius and usually better metrics. The CCPP interpretability case study is a nice idea, though it remains post hoc. The appendices list hyperparameters and limitations clearly, and the limitations section is candid.\n\nThe soft spots are real and, in places, load-bearing. The main empirical claim—\"strongest or tied-strongest primary metric on all eight datasets\"—rests entirely on Table 1, and that table cannot be reconstructed from the stated configurations. For example, HYDRA's CCPP row reports 4.8k trainable parameters. With w=43, K=6, r=6, the low-rank block alone is 2*43*6 + 36*7 = 768 parameters per block; the depth is never given, and the input/output layers add only about 200, so the reported total implies a specific depth (roughly six blocks) that is never stated. More troubling, KAN's CCPP entry reports 7.0k parameters for w=51, K=12, but a single 4->51 KAN layer already costs 4*51*13 = 2,652, a 51->1 layer adds 663, and a dense 51->51 block would cost 33.8k. No stated configuration produces 7.0k. The metrics are also internally inconsistent: Phoneme accuracy is 0.885 in Table 1 but 0.856/0.845 in Tables 3 and 5; Energy RMSE is 0.706 in Table 1 but 1.193/1.384 in Tables 3/5. All runs use one seed and one split, so the \"best on all eight\" claim is compatible with selection over many configurations.\n\nNone of this sinks the underlying idea, but it means the parameter-efficiency claim—the second half of the central thesis—is not currently supported by auditable evidence. The authors need to release code, reconcile the tables, specify depth and full model formulas, and report variance across seeds and splits.\n\nWho gets value: researchers working on KAN variants or hyperbolic representation learning will want to know about the low-rank prototype mechanism and the radius-control objective. The paper deserves a serious referee, but in its current form I would not accept; it needs a major empirical overhaul.","headline":"A clever low-rank hyperbolic KAN with a clean density proof for the full-rank case, but unreproducible parameter counts and internally inconsistent tables undermine the main empirical claim.","tokens_in":16296,"tokens_out":4461,"would_cite":false,"duration_ms":38769,"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":"This paper reports that a hyperbolic, low-rank variant of Kolmogorov-Arnold Networks matches or beats standard KANs and MLPs on all eight tabular benchmarks while using fewer trainable parameters.","keywords":["Kolmogorov-Arnold Networks","hyperbolic representation learning","Poincaré ball","low-rank prototype","parameter efficiency","tangent-space spline updates","radius control","tabular benchmarks"],"falsifier":"Train HYDRA and a full-rank Euclidean KAN of matched width on a synthetic tabular target whose hidden interactions are deliberately high-dimensional, such as a sum of many independent nonlinear pair terms that no radial coordinate can absorb. If HYDRA needs $r$ close to $d$ to match the full-rank KAN's accuracy, then the low-effective-dimension premise fails and the reported savings cannot be attributed to the hyperbolic geometry.","tokens_in":15304,"feed_emoji":"🌀","tokens_out":11601,"duration_ms":90187,"temperature":0.7,"pith_summary":"The paper proposes HYDRA, a variant of Kolmogorov-Arnold Networks that keeps KAN-style spline function learning but moves it into the tangent space of a bounded hyperbolic (Poincaré-ball) representation. Its central claim is empirical: across eight tabular benchmarks, HYDRA achieves the strongest or tied-strongest primary metric—RMSE for regression, accuracy for classification—while using fewer trainable parameters than Euclidean KAN and MLP, with an average parameter reduction of 34.9% relative to KAN and 37.1% relative to MLP. If true, this matters because KANs' main practical bottleneck is the quadratic parameter growth of dense hidden-to-hidden spline blocks, and HYDRA cuts that cost from $O(d^2K)$ to $O(dr + r^2K)$ without sacrificing accuracy. The paper also argues that the final hyperbolic radius and latent path geometry give an inspectable, physically meaningful diagnostic, illustrated on the CCPP dataset where an ambient-temperature sweep moves the representation from small-radius to large-radius and reduces predicted power by 37.36 MW.","feed_headline":"HYDRA beats or ties KAN and MLP on 8 datasets with fewer parameters","feed_subtitle":"Low-rank spline updates in a hyperbolic tangent space cut parameter counts and add an inspectable radial coordinate.","key_machinery":"The load-bearing mechanism is the low-rank prototype functional block executed in the tangent space of the Poincaré ball $\\mathbb{B}^d_c = \\{h \\in \\mathbb{R}^d : c\\|h\\|_2^2 < 1\\}$, with exponential and logarithmic maps at the origin as the fixed coordinate chart. The block writes the tangent-space KAN update as $z_{l+1} = z_l + \\alpha_l W_\\uparrow \\Phi_l(W_\\downarrow z_l)$, so the dominant cost falls from $O(d^2K)$ to $O(dr + r^2K)$; rank $r$ is a compression knob the authors choose as the smallest value that preserves the primary metric. Two complementary controls stabilize the geometry: a hard projection $\\Pi_{r_l}$ keeps every hidden state inside a per-layer radius budget, and a soft penalty $L_{\\text{rad}}$ discourages outward drift before the projection activates, preventing the near-boundary distance and gradient amplification that otherwise lets a model win by pushing samples outward instead of learning smooth spline responses. The final hyperbolic radius and latent path length are then available as interpretability diagnostics.","core_discovery":"HYDRA's core claim is that a KAN's expensive hidden-to-hidden functional map can be replaced by a low-rank prototype update inside a hyperbolic representation without giving up predictive accuracy. Each block computes $z_l = \\log_c^0(h_l)$ in tangent coordinates, projects down with $W_\\downarrow \\in \\mathbb{R}^{r \\times d}$, applies a spline block $\\Phi_l$ to the $r$ prototype coordinates, lifts with $W_\\uparrow$, adds a scaled residual, and re-embeds under a radius projection $\\Pi_{r_l}$. The authors report this architecture achieves the strongest or tied-strongest primary metric on all eight datasets in Table 1—for instance, reducing Parkinsons Telemonitoring RMSE from 4.424 to 3.534 while cutting parameters from 2.4k to 1.4k—and they support the decomposition claim with ablations in which low-rank models use a median of 33.8% of full-rank HYDRA's parameters, and radius-constrained models reduce mean radius while improving the primary metric. The universal-approximation appendix shows that when $r \\ge d$ the prototype block recovers the full tangent-space KAN update, so HYDRA is a parameter-efficient subfamily of a dense universal class.","pith_inferences":["Beyond the paper: the low-rank prototype update is geometry-agnostic in its parameter counting, so the same bottleneck could be tested in Euclidean or spherical latent spaces; that comparison would isolate whether the gains come from curvature itself or from the radius constraint and rank compression.","Beyond the paper: the final hyperbolic radius could be tested as a lightweight difficulty or out-of-distribution signal—if radius consistently tracks error or low confidence, the interpretability diagnostic becomes a usable uncertainty estimate rather than a post hoc visualization.","Beyond the paper: the effective-rank assumption implies a measurable compressibility test on any new dataset—estimate the rank of the tangent-space update Jacobian on validation data and check whether the selected $r$ tracks it; datasets with high effective rank should demand larger $r$."],"forward_implications":["The dominant hidden-to-hidden cost of KAN-style networks drops from $O(d^2K)$ to $O(dr + r^2K)$, so wider hidden layers become affordable at a fixed spline budget—an architectural saving, not a pruning artifact.","Radius control is more than numerical safeguarding: ablations show constrained models reach smaller mean radii and better primary metrics than unconstrained models, indicating the representation itself changes.","The final hyperbolic radius and latent path length can serve as HYDRA-specific interpretability diagnostics; the CCPP case study links an ambient-temperature sweep to a 37.36 MW drop in predicted power with the expected radius and SHAP direction.","Universal approximation is retained within the low-rank family: the appendix shows $r \\ge d$ recovers the full tangent-space KAN update and HYDRA networks are dense in $C(K,\\mathbb{R}^m)$.","Across the eight datasets the average parameter count is 34.9% below Euclidean KAN and 37.1% below MLP while the primary metric is competitive or better, so the method, if correct, gives interpretable functional learning at a budget close to dense neural networks."],"supporting_citations":[{"why":"It defines the original KAN architecture whose dense hidden-to-hidden spline map is the $O(d^2K)$ baseline HYDRA compresses.","marker":"Liu et al. 2025"},{"why":"It introduces Poincaré-ball embeddings and motivates hyperbolic representation learning for compact latent structure.","marker":"Nickel and Kiela 2017"},{"why":"It supplies the exponential and logarithmic maps and the tangent-space formalism HYDRA uses for its functional updates.","marker":"Ganea, Bécigneul, and Hofmann 2018b"},{"why":"It is the OpenML platform from which the eight benchmark datasets are taken.","marker":"Vanschoren et al. 2013"},{"why":"It supplies SHAP values, the reference interpretability tool used alongside HYDRA's radial diagnostics in the case study.","marker":"Lundberg and Lee 2017"},{"why":"It is the parameter-reduced Chebyshev KAN baseline in Table 1 against which HYDRA's primary metric is compared.","marker":"Sidharth et al. 2024"}],"fun_headline_variants":["HYDRA: low-rank hyperbolic KAN, same accuracy, fewer parameters","HYDRA: hyperbolic KAN with low-rank prototypes, fewer params","HYDRA: hyperbolic low-rank KAN matches or beats on 8 datasets","HYDRA: hyperbolic space cuts KAN's parameter bloat"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the premise that on these tabular tasks the hidden functional update is effectively low-dimensional, so a small prototype rank $r$ preserves what a full-rank KAN block would contribute—a premise supported only by post hoc rank-selection ablations, not by a prior characterization or theory of when low effective dimension holds.","fun_headline_variants_meta":{"raw":{"variants":["HYDRA: low-rank hyperbolic KAN, same accuracy, fewer parameters","HYDRA: hyperbolic KAN with low-rank prototypes, fewer params","HYDRA: hyperbolic low-rank KAN matches or beats on 8 datasets","HYDRA: hyperbolic space cuts KAN's parameter bloat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"total_tokens":2802,"prompt_tokens":990,"completion_tokens":1812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":1731}},"tokens_in":606,"tokens_out":1812,"duration_ms":13157,"temperature":1.0,"reasoning_tokens":1731,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:12:57.288189+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train HYDRA and a full-rank Euclidean KAN of matched width on a synthetic tabular target whose hidden interactions are deliberately high-dimensional, such as a sum of many independent nonlinear pair terms that no radial coordinate can absorb. If HYDRA needs $r$ close to $d$ to match the full-rank KAN's accuracy, then the low-effective-dimension premise fails and the reported savings cannot be attributed to the hyperbolic geometry.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces Poincaré-ball embeddings and motivates hyperbolic representation learning for compact latent structure."}],"review_version":1}