acceptodds
Under review as a conference paper at ICLR 2027

Grow or Prune? Constructive Neuroevolution of Kolmogorov–Arnold Networks

Abstract

Kolmogorov–Arnold networks (KANs) are usually trained as fixed, over-parameterized graphs, then pruned and converted to equations. We study the constructive alternative: grow the graph from a minimal network. Constructive KAN (C-KAN) evolves typed, spline-edge graphs with short Lamarckian adaptation. Identity-oriented growth, fan-in normalization, amplitude-corrected Safe Mutation and behavioral speciation address mismatches between NEAT and function-valued edges; several species supply candidates for symbolic conversion. On 30 clean SRSD equations, C-KAN-GSR returns 12 high-accuracy equations, compared with none for two fixed-template PyKAN pipelines, and its delivered equations are more often valid and smaller. It also improves black-box SRBench predictive AUC over both templates. Yet neither KAN pipeline recovers an exact skeleton: PySR and KAN-SR recover 43.3% and 15.8%, respectively. A post-hoc constant refit raises PyKAN-1 above C-KAN in high-accuracy rate, but C-KAN's expressions remain smaller and structurally closer to the targets. Stage records locate C-KAN's main bottleneck in graph search: 17 of 30 equations lack an accurate prepared network; conversion and export cause further losses. Small ablations associate the KAN-specific mechanisms with higher validity and lower measured cost, without establishing a recovery gain. Constructive growth makes sparse KANs easier to export, but graph discovery alone does not solve equation recovery.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Kolmogorov–Arnold networks, neuroevolution, neural architecture search, symbolic regression, constructive graph growth

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sCAyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

C-KAN grows Kolmogorov–Arnold networks by neuroevolution, which is a replacement of training a fixed template and pruning it. Individuals are typed sparse graphs with spline edges. A NEAT-style loop applies five Adam steps per generation and writes the adapted weights back (Lamarckian inheritance). The authors name five problems that arise when NEAT meets spline edges and add a fix for each: growth at the identity function, gradual fan-in normalization with geometric-mean multiplicative nodes (Proposition 1 bounds activation scale), crossover inside the fitter parent’s template, safe mutation with the spline amplitude divided out, and a bounded behavioral distance for speciation. For symbolic regression, five champions spread across species are pruned by replacing edges with their mean output. They are then converted per edge (“vanilla”) or by greedy in-context symbolic regression (GSR), and the simplest valid equation within a fidelity band is returned.

Experiments use SRBench and 30 clean SRSD equations. Baselines are PyKAN with one or two inner layers (vanilla extraction only, sparsity weight 0), KAN-SR, PySR, Operon and TPSR. C-KAN-GSR returns 12 of 30 equations with test R^2 above 0.999, where neither PyKAN template returns any, and its black-box AUC is 0.332 against 0.249 and 0.149. No KAN pipeline recovers an exact skeleton, while PySR recovers 43.3% and KAN-SR 15.8%. A constant refit added after seeing the results lifts PyKAN-1’s high-accuracy rate to 54%, above C-KAN-GSR’s 45%, though C-KAN’s expressions stay smaller (median size ratio 2.0 against 7.4). A stage breakdown places 17 of 30 losses in the search, 3 in conversion and 7 in export. Ablations on 12 and 6 equations show zero recovery in every variant.

Critical strengths
  • I think, the main strength is the stage-level diagnosis that the paper reports even where it cuts against the headline. Tables 2 and 21 trace each equation to the first stage where it is lost. The finding that the dense template fits more equations at first (22 against 9) but loses them later, while the grown graph keeps what it finds, points to where KAN pipelines break. That is useful beyond this method, even though the comparison is not yet like for like (see the first weakness below).

  • The second strength is the good scoping together with a concrete account of why plain NEAT fails on spline edges. The authors name five specific mismatches and mark each mechanism as new, adapted or borrowed (Table 4). The amplitude correction to Safe Mutation addresses a real problem for low-amplitude splines. The paper labels its comparisons as exploratory and its clean-up analyses as post hoc, and it reports the PyKAN refit divergence and the failure on the 50-input dataset. I did not check the proofs.

Critical weaknesses
  • The first weakness is that the central claim, that growing beats pruning, rests on a weak, partly diverged and unevenly instrumented baseline, and the paper’s own refit result undercuts it. PyKAN ran with sparsity weight 0, although that penalty is part of the standard recipe before pruning. Only vanilla extraction was completed, so C-KAN-GSR is compared with a pipeline that never received GSR. PyKAN’s post-pruning refit diverged in 85 of 150 Track B runs, and PyKAN-2’s median expression has 1,178 nodes, which suggests its pruning did not work as intended. The stage comparison in Table 2 is also not like for like, because C-KAN prunes and refits in one step while PyKAN’s symbolic model is not logged before its refit. The roster was revised after results were seen, so every test is exploratory. After Holm correction, no Track A paired comparison passes 0.05, and the NED comparison with PyKAN is not detectable either. With constants refit, PyKAN-1 beats C-KAN on accuracy (54% against 45%), and both recover 5.8% of skeletons. What remains is smaller expressions, but C-KAN’s own pruning and selection rules also favour small size, so the size advantage is not tied to growth. GSR and vanilla outputs also share one evolution, with GSR run first, so the GSR gain is not isolated.

  • The method misses its stated task, and the ablations cannot link results to the proposed mechanisms. No exact skeleton is recovered, and 17 of 30 equations have no accurate prepared network, so the main bottleneck is the very search the paper proposes. The mechanism ablations use 12 and 6 equations where recovery is zero everywhere, so only validity and runtime can differ, and even those comparisons are descriptive because hardware and budgets differ. No control separates growth from minimal networks, which is the paper’s central design choice, from evolutionary search over fixed topologies at equal compute. Black-box AUC and expression size also stay behind PySR and Operon, and 10 of C-KAN’s 16 invalid Track A outcomes come from one 50-input dataset.

Questions for authors

1 - What are the wall-clock and compute budgets for each method, including PyKAN’s training, pruning and refit, on the same hardware? C-KAN uses 100 generations of 150 individuals within one hour on one CPU thread, so how do PyKAN, PySR and KAN-SR compare on that budget?

2 - How does PyKAN do with a tuned sparsity weight, with its refit divergence fixed, and with GSR extraction applied to its networks, and does the vanilla-versus-vanilla gap survive that?

3 - What happens if C-KAN’s own pruning (mean replacement with the tolerance rule), refit and selection are applied to the trained PyKAN-1 network, and can PyKAN’s symbolic model be logged before its refit so that the stage columns in Table 2 match?

4- Can you add an equal-budget control that runs the same evolutionary loop over fixed-size topologies, without growth from minimal networks? Without it, I cannot attribute the result to constructive growth.

5 - Please report per-seed high-accuracy and recovery rates with intervals, not only best-of-five by validation loss.

5 comments
AuthorsOpenReview ↗

Response to Reviewer sCAy

Thank you for the detailed experimental and presentation concerns. We have expanded the collection to 24 SRBench datasets × 30 seeds and all 120 SRSD equations × 10 seeds, reporting unresolved execution cases explicitly. We are strengthening the comparisons and rewriting the methodology/results to clarify stage definitions, evidence and limitations.

1. Budgets and runtime reporting. Each main pipeline has a 3,600-second training-to-equation allowance, plus 120 seconds for evaluation. C-KAN uses one CPU thread; PyKAN uses CUDA with one host thread.

The experiments were collected across heterogeneous machines with concurrent workers. We will report hardware configurations, recorded stage timings, hardware-specific medians/IQRs and timeout rates, with concurrency disclosed where known. Published SRBench records will be distinguished from locally measured runs.

These measurements describe computational costs under our collection conditions; they do not establish equal compute or hardware-independent speedups. We will make this limitation explicit and avoid pooled cross-machine runtime rankings. Would this budget/resource disclosure and machine-specific timing analysis address your concern? Clarification on whether your principal concern is resource allocation or comparative execution speed would help us focus the reporting.

2. Stronger PyKAN and matched extraction. We added staged sparsity-regularized training, pruning, grid refinement and refitting, with a two-depth regularization pilot (λ = 0, .0001, .001, .01; selected .01), and both vanilla/GSR extraction. The original 85/150 divergence count concerns the unregularized recipe; failures of the new recipe will be reported separately, distinguishing numerical outcomes from execution failures.

The vanilla-versus-vanilla advantage survives on Track A but reverses on Track B: original validation-selected high-accuracy counts are 37/120 for C-KAN versus 54/120 for tuned PyKAN-1. For GSR, Track A’s 12 complete first-principles datasets favor C-KAN throughout: failure-aware median dataset R² .802 versus .238 (exploratory Holm-adjusted p = .00244).

After common Track B backward elimination, rounding and constant refitting, C-KAN-GSR versus PyKAN-1-GSR gives:

  • High accuracy (R² > .999): 54/120 versus 73/120.
  • Exact recovery: 11/120 versus 6/120.
  • Mean scored NED: .832 versus .909.
  • Conditional median expression size: 14 versus 28 nodes.

This is a prediction–structure trade-off, not general superiority. Original exact recovery remains zero for both KAN pipelines. Cleanup adds computation and remains post hoc; fixing the original selected seed narrows recovery to 8 versus 7. Two PyKAN-GSR execution-failure cells remain under verification, making its figures provisional.

3. C-KAN-style pruning and matched stages. The requested control is implemented: one extended-training PyKAN checkpoint feeds native versus mean-preserving C-KAN-style pruning, each followed by vanilla/GSR extraction. Dataset-specific training targets derive from recorded C-KAN evolution times; its 5,400-second per-output cap is deliberately more generous, not equal compute.

New logging records pruning/refitting stages and symbolic conversion before final refitting. We distinguish the numerical champion from the source of the reported equation. This control is unfinished, so no pruning conclusion is claimed. We will also apply common post-hoc cleanup to Track A and the ablations.

4. Fixed-topology evolution and attribution. A ten-hidden-node control retains the evolutionary/adaptation/extraction settings with node/link growth disabled and the main time cap. Results are incomplete, so compactness remains a pipeline-level finding rather than proof of a causal growth benefit.

Further mechanism/candidate-count ablations are completed. The full SM/behavior configuration yields 28/60 fully evaluable GSR outputs versus 19–24/60 for removals, but recovery remains zero. Hardware/timing qualifications and the 50-input Track A scalability failure remain explicit.

5. Individual-seed results. Original Track B high-accuracy rates across individual runs, with 95% intervals, are:

  • C-KAN GSR: 28.8% [21.8–35.9%].
  • PyKAN-1 GSR: 23.2% [18.3–28.3%].
  • C-KAN Vanilla: 19.8% [13.6–26.4%].
  • PyKAN-1 Vanilla: 28.6% [22.3–35.2%].

Intervals resample whole equations, keeping their seeds together. These complement validation-selected results. Revised-roster tests and cleanup remain explicitly exploratory/post hoc.

Thank you for the new experiments. Tuned PyKAN with GSR beats C-KAN on Track B accuracy (73/120 against 54/120), so it looks like the result is a trade-off with expression size. The pruning and fixed-topology controls are unfinished, so growth is still not separated from C-KAN’s own pruning and selection rules. I could not check the expanded benchmark in this period. I am keeping my score for now. The paper would be stronger as a resubmission with the finished controls and claims stated as a trade-off.

AuthorsOpenReview ↗

Extra ablation runs completed

Thank you for the clarification. We agree that the results should be presented as trade-offs, rather than general superiority over tuned PyKAN.

The fixed-topology, pruning-transfer and longer-fitting controls are now complete on the same 12-equation, five-seed panel. The following equation counts include common elimination/rounding/refitting and retain failures in the denominator:

  • Growth: Keeping C-KAN’s minimal graph fixed, with the same recovery procedure, yields 5/60 highly accurate GSR equations versus 11/60 with growth. However, the fixed minimal graph is more often fully evaluable (54/60 versus 36/60). The fixed-[10] follow-up now completes all 100 generations under a six-hour allowance, yielding 10/60 highly accurate GSR equations.
  • Pruning transfer: Normally trained tuned PyKAN with C-KAN-style pruning yields 9/60 highly accurate GSR equations versus 22/60 with native pruning. Thus transferring the pruner does not reproduce C-KAN’s delivery behavior.
  • Additional fitting: Post-pruning L-BFGS increases numerical high accuracy from 33/60 to 37/60. Cleaned symbolic accuracy improves for Vanilla (15→19/60), while GSR remains at 22/60. We are incorporating these completed controls and their limitations into the revision. They sharpen the fitting–delivery trade-off, but do not establish that growth alone explains every difference.
Reviewer_8QTHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Constructive KAN (C-KAN), a framework that progressively grows KAN computational graphs from minimal networks through neuroevolution, as an alternative to the conventional train–prune–symbolify pipeline based on fixed network architectures. To accommodate the unique properties of spline-valued edges, the authors introduce several mechanisms, including constructive graph growth, fan-in normalization, Safe Mutation, and behavioral speciation. The framework further employs symbolic conversion of multiple candidates to extract mathematical equations.

Critical strengths

Well-defined research problem. The authors identify the limitations of fixed-architecture KANs in representing complex structures and recovering symbolic expressions. Investigating constructive graph search as an alternative is a meaningful research direction.

**Well-motivated technical design. **The paper introduces several mechanisms specifically designed to address the challenges of applying NEAT to function-valued edges, rather than simply combining existing neuroevolution techniques.

Transparent experimental analysis. The authors provide detailed analyses of performance degradation at different stages, including structural search, network pruning, symbolic conversion, and final expression export. The paper also explicitly acknowledges the failure cases and limitations of the proposed approach.

More compact symbolic expressions. Compared with fixed-template PyKAN, C-KAN produces smaller expressions that are structurally closer to the target equations in several experiments.

Critical weaknesses

Limited performance on the core task. On the 30 SRSD equations, C-KAN achieves a 0% exact structural recovery rate, whereas PySR and KAN-SR achieve 43.3% and 15.8%, respectively. Given that symbolic equation recovery is a central objective of the proposed framework, this result substantially limits its practical contribution.

Limited robustness of the improvements over fixed-architecture KANs. Although C-KAN outperforms PyKAN in the originally exported results, PyKAN-1 achieves a higher high-accuracy rate after post-hoc constant refitting. This suggests that the initially reported advantage may partly arise from differences in parameter fitting and expression export procedures, rather than from the constructive graph search itself.

Concerns about baseline fairness. The fixed-template PyKAN baselines use the default sparsity settings, without sufficiently comprehensive hyperparameter tuning. Stronger comparisons under unified post-processing procedures are needed to establish the advantages of the proposed method.

Insufficient evidence from the ablation studies. Although some proposed mechanisms improve the rate of valid expression generation or reduce computational costs, the small-scale ablation studies do not demonstrate clear improvements in exact structural recovery.

Questions for authors

1, Given that structural search remains the main bottleneck, what unique advantages does constructive graph growth offer over direct symbolic expression search?

2, Does C-KAN still significantly outperform fixed-template KAN approaches when the baselines are properly tuned and evaluated under matched computational budgets and identical constant-refitting procedures?

3, Could the authors provide a more detailed analysis of the factors responsible for the zero exact structural recovery rate and conduct targeted experiments to address these limitations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8QTH

Thank you for highlighting baseline fairness, attribution and recovery as the decisive issues. We have expanded the benchmark collection to 24 SRBench datasets × 30 seeds and all 120 SRSD equations × 10 seeds, with unresolved execution cases reported explicitly. We are strengthening the baselines and clarifying the methodology and results.

1. What does constructive growth offer over direct expression search?

C-KAN learns a continuously trainable KAN graph without prescribing hidden depth, width or multiplication locations. Its closest architectural comparator is therefore fixed-template PyKAN. KAN-SR additionally uses SR-specific symmetry/separability decomposition, which C-KAN does not incorporate; direct SR methods remain important performance references.

The empirical contribution is a study of representation and extraction trade-offs, rather than superior exact recovery. C-KAN can deliver compact, accurate equations relative to fixed KAN pipelines, but original C-KAN outputs still recover zero exact skeletons across the larger clean-SRSD collection. We retain this limitation prominently.

2. Does the advantage survive stronger PyKAN, budgets and common refitting?

Not uniformly. We added staged sparsity-regularized PyKAN, a regularization pilot (λ = 0, 0.0001, 0.001, 0.01; selected λ = 0.01), and GSR alongside vanilla extraction. The revised comparisons distinguish extraction modes.

On Track A’s 12 complete first-principles datasets, C-KAN-GSR exceeds tuned PyKAN-GSR on every failure-aware dataset accuracy summary; the median across datasets is .802 versus .238 (exploratory Holm-adjusted p = .00244). Black-box run-average AUC is .482 versus .279–.349; the latter range bounds unresolved PyKAN outcomes, rather than representing a confidence interval.

On Track B, common post-hoc backward elimination, rounding and constant refitting give this current GSR comparison:

  • High accuracy (test R² > 0.999): C-KAN 54/120; PyKAN 73/120.
  • Exact recovery: C-KAN 11/120; PyKAN 6/120.
  • Mean failure-penalized NED: .832 versus .909.
  • Conditional median expression size: 14 versus 28 nodes.

Thus, selected predictive accuracy favors PyKAN, while structural scores and compactness favor C-KAN. Two PyKAN execution-failure cells remain under verification, so these GSR figures are provisional. Holding the originally selected seed fixed narrows recovery to 8 versus 7. Cleanup is reported separately because it adds computation; we will also apply it to Track A and the ablations.

The main allowance is 3,600 seconds for training/extraction per logical output, plus 120 seconds for evaluation. CPU/GPU and collection-machine differences preclude an equal-compute claim. The 5,400-second PyKAN control is an additional-budget diagnostic.

3. What explains zero recovery, and which targeted tests address it?

The evidence identifies multiple limitations. Numerical fitting can fail before extraction; downstream conversion, redundant terms and coefficients can also prevent exact structural matching. Common cleanup exposes recoverable structure, but does not establish that original graph search succeeded. We are extending numerical-versus-symbolic diagnostics and distinguishing the numerical champion from the candidate supplying the final equation.

Two targeted controls are in progress: the same evolutionary loop with growth disabled, and extended PyKAN training followed by native versus C-KAN-style pruning, each with vanilla/GSR extraction and additional stage logging. Until complete, these do not establish a causal growth advantage.

We have completed further mechanism and candidate-count ablations. The full SM/behavior configuration yields 28/60 fully evaluable GSR outcomes versus 19–24/60 for removals, but recovery remains zero. We will retain that distinction, hardware/timing qualifications, and exploratory/post-hoc labels, while adding per-seed results with equation-level uncertainty estimates.

The supplementary noise-and-dummy study originally had inconsistent cross-machine corruption. Its data controls are now corrected using identical frozen inputs. Preliminary results on 979 matched pairs show higher C-KAN-GSR predictive validity (92.2% versus 66.6%), but lower high-accuracy success (27.1% versus 32.9%); this is not a universal robustness advantage.

We are rewriting the methodology and results to clarify stage definitions, separate formal properties from demonstrated empirical benefits, and make the scope and limitations easier to follow.

Dear Reviewer,

As we are approaching the end of the discussion period, we would like to kindly ask whether you have any remaining concerns or questions that we could further address or clarify.

We also noticed that one of the reviewers has updated their assessment following the rebuttal. In light of this, we would sincerely appreciate it if you could kindly revisit our responses and reconsider your initial assessment, particularly if you feel that we have adequately addressed your concerns.

AuthorsOpenReview ↗

Final revision update: completed growth and PyKAN controls

We’ve uploaded the revised manuscript. Section 4.5 reports the controls that were still in progress in our earlier reply. Starting from the same minimal graph and using the same recovery procedure, allowing growth increased highly accurate GSR equations after common cleanup from 5 to 11 of 60, while structurally evaluable equations fell from 54 to 36. Longer fitting improved tuned PyKAN’s numerical accuracy, while its cleaned GSR high-accuracy count stayed at 22 of 60. Both KAN pipelines still recover zero exact skeletons as exported on the full clean benchmark. The revision presents these results as a trade-off between predictive accuracy and structural evaluability.

Reviewer_ndMUOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C-KAN, a constructive framework that grows the structure of Kolmogorov-Arnold Networks from a minimal graph instead of training and pruning a fixed overparameterized architecture. The authors adapt NEAT-style neuro evolution to KANs with several mechanisms for stable topology growth, functional mutation, speciation, and symbolic recovery. Experiments on SRBench and SRSD show that C-KAN produces more accurate, valid, and compact equations than the tested fixed-template PyKAN pipelines, while its exact symbolic recovery remains substantially behind direct symbolic regression methods.

Critical strengths
  1. The paper explores an alternative to the standard train-prune-symbolify pipeline by directly searching for sparse KAN structures through constructive growth.
  2. The proposed KAN-specific mechanisms address practical issues that arise when applying NEAT-style evolution to function-valued spline edges, including activation scaling, mutation stability, and behavioral speciation.
  3. The stage-level analysis is informative. It shows that constructive and destructive KANs fail at different stages: fixed-template KANs can fit more equations initially but lose accuracy during pruning and export, while C-KAN is more likely to preserve the accuracy of the structures it successfully discovers.
  4. C-KAN produces substantially smaller equations than the tested fixed-template PyKAN baselines, suggesting that constructive growth can provide a better intermediate representation for symbolic conversion.
Critical weaknesses
  1. The main limitation is that the proposed constructive search still fails at exact symbolic recovery. C-KAN recovers no exact skeleton on the clean SRSD benchmark, while PySR and KAN-SR achieve substantially higher recovery rates. The authors' own stage-level analysis further shows that graph search is the dominant source of failure.
  2. The advantage over fixed-template KANs is less clear after post-hoc constant refitting. PyKAN-1 can recover strong numerical accuracy after refitting, suggesting that part of C-KAN's reported advantage comes from the downstream recovery pipeline rather than constructive topology search itself.
  3. The ablations provide limited evidence for several proposed components. Safe Mutation and behavioral speciation mainly improve validity or computation-related measures, but do not establish improvements in accuracy or symbolic recovery.
  4. C-KAN introduces a relatively complex evolutionary pipeline with several KAN-specific mechanisms, while direct symbolic regression methods such as PySR still achieve substantially better structural recovery. This raises the question of whether the additional complexity is justified for symbolic regression.
  5. The experiments are based on a relatively limited subset of symbolic regression benchmarks, and several comparisons and diagnostic analyses are exploratory or post-hoc. A broader and more controlled evaluation would strengthen the conclusions.
Questions for authors

Listed in Weakness part

4 comments
AuthorsOpenReview ↗

Response to Reviewer ndMU

Thank you for recognizing the constructive formulation and stage-level diagnosis. We have expanded the benchmark collection to 24 SRBench datasets with 30 seeds and all 120 SRSD equations with ten seeds, reporting unresolved execution cases explicitly. We have also strengthened PyKAN and standardized post-processing to test which original conclusions persist.

1. Exact recovery and the search bottleneck. Recovery remains a central limitation: original C-KAN outputs still recover no exact skeletons in the larger clean-SRSD collection. We retain this result and the stronger direct-SR baselines prominently. We are sharpening the research question to ask what changes when KAN graphs are grown rather than trained as fixed templates and subsequently pruned. The numerical-versus-symbolic diagnostics help distinguish fitting limitations from failures in equation extraction and export.

2. Refitting and baseline fairness. We have added staged, sparsity-regularized PyKAN, a four-strength regularization pilot, and both vanilla and GSR extraction. We completed common post-hoc backward elimination, rounding and constant refitting for both pipelines. The current GSR comparison after cleanup is:

  • High accuracy (test R² > 0.999): C-KAN 54/120; PyKAN 73/120.
  • Exact recovery: C-KAN 11/120; PyKAN 6/120.
  • Mean failure-penalized NED: C-KAN .832; PyKAN .909.
  • Conditional median expression size: C-KAN 14 nodes; PyKAN 28.

Thus, tuned PyKAN retains the selected predictive-accuracy advantage, while C-KAN has better structural scores and smaller selected expressions. These figures remain provisional while two PyKAN execution-failure cells are verified. Holding the originally selected seed fixed narrows recovery to 8 versus 7, so we do not attribute the difference solely to growth. We will also apply the same cleanup to Track A and the ablation studies. Cleanup results will remain separate from original outputs, with their additional computation disclosed.

3. Ablation evidence. We agree that the component studies do not establish recovery gains. In the SM–behavioral-speciation factorial, the full configuration yields 28/60 fully evaluable GSR outcomes versus 19–24/60 for the removals, with zero recovery throughout. Additional candidate-count and mechanism ablations are completed. Applying common cleanup will test whether their differences persist after downstream processing, but no cleanup benefit is assumed in advance. Historical hardware and timing differences remain disclosed.

4. Justification and comparison scope. The contribution is an alternative way to learn sparse, continuously trainable KAN representations. KAN-SR additionally uses SR-specific decomposition strategies, including symmetry and separability, which C-KAN does not incorporate. PyKAN is the closest architectural comparator for the constructive-versus-fixed-template question within the KAN representation family. Other SR methods remain important performance references, and their superior recovery is acknowledged. Fixed-topology evolution and C-KAN-style pruning applied to PyKAN controls are in progress; until complete, we do not claim that growth uniquely causes the observed benefits.

5. Broader evidence and clearer presentation. On the 12 complete Track A first-principles datasets, C-KAN-GSR has better failure-aware dataset accuracy than tuned PyKAN-GSR throughout (exploratory Holm-adjusted p = .00244). Track B instead shows the trade-off above. We are adding individual-seed results with equation-level uncertainty estimates and preserving exploratory/post-hoc labels.

The original supplementary material included noise-and-dummy-variable experiments, but cross-machine corruption inconsistencies prevented a properly controlled comparison. We have corrected the data controls by freezing identical corrupted inputs across machines. Preliminary results on 979 matched dataset–seed pairs show higher predictive validity for C-KAN-GSR (92.2% versus 66.6%), but lower high-accuracy success (27.1% versus 32.9%). We therefore do not claim universally greater robustness.

We are reorganizing the manuscript around these questions, simplifying the methodology exposition, and clarifying stage definitions and the distinction between observed findings and causal interpretations.

Dear Reviewer,

As we are approaching the end of the discussion period, we would like to kindly ask whether you have any remaining concerns or questions that we could further address or clarify.

We also noticed that one of the reviewers has updated their assessment following the rebuttal. In light of this, we would sincerely appreciate it if you could kindly revisit our responses and reconsider your initial assessment, particularly if you feel that we have adequately addressed your concerns.

The rebuttal substantially clarifies the scope of the contribution and addresses several baseline-fairness concerns, but the central question of whether constructive growth itself is responsible for improved symbolic structure remains unresolved.

AuthorsOpenReview ↗

Reply to ndMU

Thank you for raising this point. We completed two no-growth controls on 12 clean SRSD equations with five matched seeds each. The first starts from the same minimal graph as C-KAN and disables only node and link additions, keeping the training and extraction rules the same. Growth improved failure-aware test R2R^2 with both extraction backends, before and after common cleanup (adjusted p≤0.042p \leq 0.042). After GSR and cleanup, growth yielded 11/60 highly accurate outputs versus 5/60. The minimal control yielded more fully evaluable equations (54/60 versus 36/60) and lower mean normalized edit distance (0.845 versus 0.938). Both recovered zero exact skeletons among the 12 validation-selected equations.

The second control starts with a fixed ten-node hidden layer (matching PyKAN-1 network). All 60 fits completed 100 generations under a six-hour allowance. After GSR and cleanup, it produced 10/60 highly accurate and 10/60 fully evaluable outputs, compared with 11/60 and 36/60 for growing C-KAN.

Together, these controls show that allowing growth improves predictive performance from the same minimal graph, while structural outcomes depend on the fixed topology used for comparison. We revised the paper to leave open whether growth itself explains the structural differences between C-KAN and PyKAN.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.