acceptodds
Under review as a conference paper at ICLR 2027

RuleKAN: Addressing the Symbolic Expressivity Gap in KANs for Symbolic Regression

Abstract

Symbolic regression tries to turn data into analytic equations that carry useful meaning. Kolmogorov–Arnold Networks (KANs) are attractive for this purpose because their learned one-dimensional edge functions can be replaced by analytic formulas. The same flexibility creates a problem: each KAN edge can absorb structure that a scientist may want the equation to expose. For example, a single learned curve can approximate , and a two-branch fuzzy-rule can be reproduced without displaying its gate and branches. We study which symbolic capacities are needed to recover those structures and more, introducing two related method families: RuleKAN trains a KAN to propose interacting variables and then searches explicit sums of products of analytic factors; RuleSISP searches the same language without the learned-support restriction. Bounded extensions of these methods add whole-expression powers, ratios, and one additional composition level. Across analytic tasks spanning products, fuzzy rules, powers and ratios, nested functions, canonical KAN targets, and physics equations, the RuleKAN family performs best overall (NRMSE ; mean rank ) against baseline methods like PySR, SR-KAN, Operon, etc. It has significantly lower predictive error than Operon, SR-KAN, and the other six KAN-based baselines (FDR-adjusted ), and wins on about 66% of tasks against PySR. Notably, on fuzzy rules, RuleSISP recovers the generating structure significantly better than PySR despite its competitive predictive error (mean expanded-rule F1 vs. ; ). The results show that symbolic recovery is not one search problem, separating distinct bottlenecks in factorization, interaction coverage, whole-expression operations, symbolic depth, and recovery of the intended structure.

Primary area:
interpretability and explainable AI
Keywords:
symbolic regression, Kolmogorov–Arnold networks, interpretable machine learning, neuro-symbolic learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eaZqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KAN-based symbolic regression methods replace each learned KAN edge with an analytic function, but this per-edge extraction can hide structure, such as a product e^x sin x absorbed into one edge or the gate and branches of a fuzzy if–then–else rule. This paper represents expressions as sums of products of analytic factors, each applying one primitive to an affine function of a single input, with bounded extensions for powers, ratios, and one level of composition. RuleKAN trains a shallow KAN-like model only to find which variables occur together, selects each factor's primitive by a gradient-based soft search followed by a discrete refit, and assembles the expression greedily term by term. RuleSISP skips the KAN and enumerates all variable combinations by brute force, which avoids missing interactions at a cost that grows quickly with the number of inputs. On 30 noise-free analytic tasks, the best variant has the lowest median error and best mean rank among 23 methods, with significantly lower error than 13 of 14 baselines (not PySR), and RuleSISP exactly recovers 8 of 21 fuzzy-rule structures versus none for PySR.

Critical strengths

The key strength is a valid improvement over KAN-based extraction: its grammar expresses factor products, interactions, and fuzzy rules that per-edge replacement cannot, and it clearly outperforms all KAN-based baselines. Its search achieves the lowest median error among 23 methods and is competitive with strong baselines such as PySR.

Critical weaknesses

The main weakness is that the KAN's contribution is unclear: it only selects variable sets, its learned functions are discarded, and RuleSISP, which uses no KAN, recovers fuzzy-rule structure better. The efficiency advantage of RuleKAN over RuleSISP is also not shown, since no runtime is reported and all tasks have at most six inputs, where brute-force enumeration is cheap.

Questions for authors
  1. Is there evidence that the representation ability of the KAN is actually used, given that only its variable sets are kept? How is the KAN better than a symbolic network or a standard interaction detector for this role?

  2. What is the measured efficiency advantage of RuleKAN over RuleSISP? Does RuleKAN scale to more complex cases, such as many more inputs, where RuleSISP is not viable?

  3. The advantage over PySR appears to be mostly in expression form rather than error. Are there concrete examples of recovered expressions showing this advantage?

3 comments

Q1: Contribution of the KAN representation

Stage 1 uses the KAN's learned one-dimensional functions to identify variable supports, but does not transfer those functions to the final equation. A support is the set of inputs in a product term. Stage 2 fits analytic factors to observed targets. For example, a KAN approximation of exsin⁡(x)e^x\sin(x) can identify support {x}\{x\}, while Stage 2 recovers the exponential and sine factors.

To test whether KAN-derived supports are preferable to those from a standard interaction detector, we also compared them (for this rebuttal) with multilayer perceptron neural interaction detection (MLP-NID) [1] on all 30 tasks (three seeds each). Both configurations use the same PowerRuleKAN-Comp fitting procedure; only the source of supports differs: either KAN or MLP-NID. KAN achieves lower NRMSE on 17/30 tasks and median task-level NRMSE 0.001071 vs 0.001664 for MLP-NID (lower is better). The results differ by task family: KAN performs better on physics equations (0.000325 vs 0.005318), while MLP-NID performs better on nested expressions (0.0000364 vs 0.002157). On seven fuzzy-rule tasks, KAN has higher mean expanded-rule F1 (0.357 versus 0.249; higher is better), although both achieve 5/21 exact recoveries.

KAN-derived supports can help, but the advantage depends on target structure. This is consistent with the no-free-lunch theorems [2], which rule out a universally superior algorithm under their assumptions. Different support detectors may therefore suit different symbolic regression problems.

We’ll add these considerations and results to the paper.

Q2: Efficiency and scaling

Yes. RuleKAN can scale more efficiently than RuleSISP as inputs grow. For dd inputs and fixed factor order qq, RuleSISP's Stage 2 search space grows as O(dq)O(d^q), while RuleKAN restricts the search to KAN-derived supports (Section 4; Appendix C). With ten analytic primitives, ratios of two flat products yield 3.38 billion possibilities at 6 inputs (q=3q=3) and 8.71 quintillion at 50 inputs (q=4q=4), excluding deeper composition. Neither method exhaustively evaluates these expressions: RuleSISP screens primitives in stages, while RuleKAN additionally restricts variable supports.

On a 50-input SRBench dataset (650_fri_c0_500_50), with over eight times the maximum input dimension of our original 30-task benchmark (Appendix A, Table 2), RuleKAN-Comp and PowerRuleKAN-Comp finish in 206 and 432 seconds, respectively, while RuleSISP and RuleSISP-Comp time out after 2,400 seconds. Further details on the SRBench experiments are provided in our response to Reviewer uLBi, Q2. These results support the computational advantage, although the measured runtimes depend on our research implementations; optimized implementations could reduce absolute runtime without changing the asymptotic complexity.

We’ll add these considerations and results to the paper.

Q3: Examples of recovered expressions

Yes. On the fuzzy-tree task, RuleKAN recovers the generating gate-and-branch structure in two of three seeds. With coefficients rounded for readability, the recovered expression is:

0.60(1−x0)(1−x1)cos⁡(1.7x2)+0.90(1−x0)x1exp⁡(−0.80x3)+0.50x0tanh⁡(1.9x4).0.60(1-x_0)(1-x_1)\cos(1.7x_2)+0.90(1-x_0)x_1\exp(-0.80x_3)+0.50x_0\tanh(1.9x_4).

Its median test NRMSE is 9.13×10−89.13\times10^{-8}, and median expanded-rule F1 is 1.0, indicating agreement with the annotated rule structure.

For the same task, PySR instead returns the following expression in seed 1:

tanh⁡(x4exp⁡(x0))3.02303+1exp⁡(−x1−cos⁡(x2))+exp⁡(x3+exp⁡(x0)).\frac{\tanh(x_4\exp(x_0))}{3.02303}+\frac{1}{\exp(-x_1-\cos(x_2))+\exp(x_3+\exp(x_0))}.

Its test NRMSE is 0.318 and expanded-rule F1 is 0: it does not expose the complementary gates and branch functions although it could. Across all seven fuzzy-rule tasks, RuleSISP achieves significantly higher mean expanded-rule F1 0.446 and 8/21 exact recoveries, vs 0 and 0/21 for PySR (Figure 4). As detailed in our response to Reviewer s4YD, Q1, adding hard if/else and soft-blend operators to PySR improved predictive accuracy but yielded no exact structural recoveries.

Regarding predictive accuracy and the advantage over PySR, PowerRuleKAN-Comp has lower test error than PySR on 20/30 original analytical tasks. In new rebuttal experiments (see our response to Reviewer uLBi, Q2), PowerRuleKAN-Comp also achieves lower error on 63/100 AI Feynman equations, while RuleKAN-Comp outperforms PySR on 20/24 SRBench 2025 datasets using standard setup configurations for those benchmarks. Thus, the fuzzy-rule results demonstrate the advantage in recovering explicit symbolic structure, while the new benchmarks provide further evidence of competitive predictive accuracy.

We’ll add these considerations and results to the paper.

References

  1. Tsang, Cheng, and Liu (2018). Detecting Statistical Interactions from Neural Network Weights. https://arxiv.org/abs/1705.04977
  2. Wolpert and Macready (1997). No Free Lunch Theorems for Optimization.

For clarity, "0.60(1−x0)(1−x1)cos⁡(1.7x2)+0.90(1−x0)x1exp⁡(−0.80x3)+0.50x0tanh⁡(1.9x4)0.60(1-x_0)(1-x_1)\cos(1.7x_2)+0.90(1-x_0)x_1\exp(-0.80x_3)+0.50x_0\tanh(1.9x_4)" is the expanded sum-of-products form of the nested fuzzy tree. Factoring out (1−x0)(1-x_0) from the first two terms reveals the nested gating hierarchy (x0x_0 followed by x1x_1).

Thank you for your response. The response has addressed most of my concerns except that the advantage of KAN-derived supports over a standard detector is small. I have raised my score accordingly.

Reviewer_uLBiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RuleKAN, a structured symbolic extraction framework for symbolic regression, aiming to address the limitations of conventional KANs in explicitly recovering multiplicative factors, variable interactions, and complex expression structures through edge-wise symbolic extraction. Specifically, RuleKAN first trains and prunes a KAN-based numerical model to identify candidate variable interactions. It then constructs a Sum-of-Products symbolic search space over these interactions and recovers analytical expressions through differentiable function mixtures, discrete selection, and parameter refitting. In addition, the authors introduce RuleSISP, which performs symbolic search without KAN-derived interaction constraints.

Critical strengths
  1. The paper addresses a meaningful research problem. The discrepancy between numerical fitting accuracy and symbolic structure recovery in conventional KAN-based symbolic extraction is an important issue. Recovering scientifically interpretable analytical structures while maintaining predictive accuracy is a worthwhile research direction in symbolic regression.
  2. The structured symbolic search framework is reasonably motivated. RuleKAN separates variable interaction identification from analytical function selection and employs an explicit Sum-of-Products representation to recover multiplicative relationships among analytical factors. This staged design provides a clear way to address different limitations in symbolic expressivity.
Critical weaknesses
  1. The evaluation on standard symbolic regression benchmarks is limited. The experimental evaluation primarily relies on specially designed synthetic tasks, without sufficient validation on established symbolic regression benchmarks such as SRBench and AI Feynman. Additional experiments on these benchmarks would help better assess the generalization ability and practical applicability of the proposed method.
  2. The symbolic expressiveness of the proposed method remains constrained by predefined expression structures. The framework relies on a predefined Sum-of-Products representation, along with extensions for powers and function compositions. As a result, it may struggle to recover the underlying symbolic structures of equations that fall outside the predefined search space. The current experiments do not sufficiently evaluate such cases, leaving the method's applicability to more complex and previously unseen expression structures unclear.
  3. The effectiveness of the KAN-guided variable interaction identification in Stage 1 is not sufficiently demonstrated. Stage 1 is designed to identify variable interactions and reduce the subsequent symbolic search space. However, the experimental results show that RuleSISP, which does not rely on Stage 1, achieves better predictive accuracy and structural recovery on several tasks. Moreover, the ablation study indicates that iterative pruning in Stage 1 provides little improvement in overall predictive error. Although the authors provide a theoretical analysis of search-space complexity, a systematic evaluation of interaction identification accuracy and practical computational efficiency is missing. Therefore, the current results do not sufficiently demonstrate the benefits or necessity of KAN-guided interaction identification, particularly for more complex equations.
Questions for authors
  1. How does RuleKAN handle target equations that cannot be represented by the predefined Sum-of-Products, Power, and Composition grammars? Is there a mechanism to adaptively expand the symbolic search space based on search feedback?
  2. The evaluation mainly focuses on 30 analytical tasks whose structures closely align with the proposed symbolic grammar. Could the authors further evaluate the method on established symbolic regression benchmarks such as SRBench or AI Feynman?
  3. Could the authors provide further evidence of the contribution of Stage 1? Specifically, it would be useful to report the precision and recall of the learned variable supports against the ground-truth interaction structures, and to compare RuleKAN and RuleSISP in terms of predictive accuracy, structural recovery, and actual runtime.
2 comments

Q1: Handling equations outside the predefined grammar

Yes. As with other symbolic regression methods, RuleKAN's vocabulary of analytic functions can be extended with additional functions (Section 2; Appendix A). Its symbolic search space can also be expanded by increasing the maximum number of factors per product, allowing additional powers and terms, and enabling bounded composition (Appendix D, Table 5). RuleKAN deliberately favours sums of products of analytic factors, with powers, ratios, and shallow composition. This inductive bias helps recover explicit factors and fuzzy-rule branches, although equations outside the configured search space may not be recovered exactly. Under their assumptions, the no-free-lunch theorems [1] rule out a single search method outperforming all others across all problems.

Our frameworks also include mechanisms that expand the symbolic search space in response to validation error. Specifically, RuleKAN Adaptive reconsiders variable supports discarded during Stage 1 pruning when the current symbolic model leaves a validation residual, accepting additions only if validation error improves. RuleKAN also uses bounded support rescue to test selected subsets and combinations of KAN-derived supports. Product-partition and recursive-partition rescue explore more complex fuzzy-rule branches. Finally, the -Comp variants can introduce an additional level of analytic composition when the existing model has insufficient validation accuracy. These mechanisms are described in Section 2 and Appendices D, F, and I.

Thus, RuleKAN can expand its search space through both configurable hyperparameters and validation-guided mechanisms. The latter are bounded and do not automatically discover arbitrary new operators or unlimited composition depth.

Q2: Evaluation on established benchmarks

Yes. For this rebuttal, we extended the evaluation to all 100 standard AI Feynman equations and all 24 SRBench 2025 datasets (12 black-box and 12 phenomenological/first-principles datasets), with three seeds per task. Since AI Feynman contains over four times as many tasks, we limited that comparison to the three methods (PowerRuleKAN-Comp, RuleSISP-Comp, and PySR, i.e., the best variants and baseline) to keep computational costs manageable, while evaluating six methods (RuleKAN, RuleKAN-Comp, PowerRuleKAN-Comp, RuleSISP, RuleSISP-Comp, and PySR) on SRBench.

On AI Feynman, PowerRuleKAN-Comp (the best here) achieved lower test error than PySR on 63/100 equations, a better mean rank (1.80 versus 2.22 among three methods), and lower median task-level test NRMSE (0.00190 versus 0.0216; lower is better).

On SRBench 2025, RuleKAN-Comp (the best here) achieved mean rank 2.85 versus 4.42 for PySR, median task-level test NRMSE 0.501 versus 0.685, and lower error on 20/24 datasets.

These results show competitive predictive performance on established physics equations and heterogeneous regression datasets.

Q3: Contribution of Stage 1

We measured Stage 1 support precision and recall on all 30 analytical tasks, with three seeds each. Precision is the proportion of proposed supports matching the generating equation; recall is the proportion of required supports identified. Stage 1 achieves mean task-median precision/recall of 0.440/0.639. After bounded support expansion, which adds candidates derived from the KAN without using the generating equations (Section 2 and Appendix F), these scores increase to 0.481/0.983. Thus, Stage 1 identifies many required supports, while the expanded candidate bank covers nearly all of them. The latter measures candidate coverage, not Stage 1 detection alone.

In terms of predictive accuracy, PowerRuleKAN-Comp outperforms the best support-independent variant, RuleSISP-Comp (0.00107 versus 0.00293 median test NRMSE; Appendix Table 7). The matched RuleKAN-Comp/RuleSISP-Comp comparison also favours KAN guidance (0.00203 versus 0.00293). However, RuleSISP achieves better structural recovery on fuzzy rules (mean expanded-rule F1 0.446 versus 0.286, with 8/21 versus 5/21 exact recoveries; Figure 4). The benefit therefore depends on support quality and the subsequent search and recovery mechanisms.

Computationally, Stage 1 restricts search to learned supports, reducing RuleSISP's O(dq)O(d^q) candidate space when few supports are retained for dd inputs and maximum factor order qq (Section 4; Appendix C). On a 50-input SRBench dataset, RuleKAN-Comp and PowerRuleKAN-Comp finish in 206 and 432 seconds, while RuleSISP and RuleSISP-Comp time out after 2400 seconds. These measurements support the complexity analysis, although the implementations are research prototypes; optimized implementations could reduce absolute runtime without changing the theoretical scaling.

References

  1. Wolpert and Macready (1997). No Free Lunch Theorems for Optimization. https://doi.org/10.1109/4235.585893

Thank you for your response. The additional experiments have addressed my concerns. I am willing to raise my score to 3.

Reviewer_s4YDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a symbolic extraction workflow for KANs and compare a set of methods using this workflow to conventional SR approaches. Their first main method is RuleKAN, which first identifies relevant variable interactions, then turns those groups into equations, and optionally considers adding more expressive operations on top of these candidates (e.g., ratios, composition, integer powers). Their second, RuleSISP works essentially the same, but instead just enumerates every variable multiset instead of starting from only those identified from the first stage of RuleKAN. They test 9 variants of their proposed method and 14 existing SR methods on a noise-free dataset generated from 30 selected equations using three seeds. Their results (Figure 3) find that in terms of mean rank by NRMSE, their methods perform well relative to the baselines.Their analysis also looks at structure discovery and what parts of the proposed pipeline seem to make a difference.

Critical strengths
  1. The writing is generally clear and reads well.
  2. The problem being examined is well stated and well motivated.
Critical weaknesses
  1. It is strange to implement these proposed symbolic extraction methods for the KAN and compare them to running existing SR directly on the data. Why not (also) try the existing SR methods for symbolic extraction from the KAN?
    1. The SR methods also don’t include some natural operators. For example, their method gets support for branching for these “fuzzy” examples but the SR methods (e.g., PySR) support including this as an operator.
    2. I suspect that this may work better and would be a more fair comparison.
  2. The benefit of the KAN guidance is step unclear. This is maybe fine as a negative result, but it does suggest further still that a more open-ended SR on the KAN would be the more interesting comparison, rather than RuleKAN vs RuleSISP.

General Comments

  • Could be nice to mention that GSR stands for Greedy in-context Symbolic Regression and GSM stands for Gated Matching Pursuit around line 119.
Questions for authors
  1. Did the authors consider using existing SR tools (e.g., PySR, SINDy, Operon) to do the step of symbolic extraction from the KAN (probably controlling for compute) and comparing that to their RuleKAN/RuleSISP approach?
    1. I mention it because many of them already support searching for structures in some of the motivating examples. For example, PySR supports a branching if/else operator.
    2. If the motivation is to improve symbolic extraction for KANs, this may be a better way towards that.
  2. Does the KAN guidance step really matter? Can a clear conclusion be drawn from these experiments?
2 comments

Q1: Using existing symbolic regression tools for KAN extraction

Yes. For this rebuttal, we tested PySR [1] as a KAN-extraction alternative by fitting symbolic expressions to predictions from a Stage 1 numerical KAN rather than to observed target values. We evaluated it on all 30 analytical tasks with three seeds each (90 runs), using the benchmark data splits. This tests whether an existing symbolic regression tool can replace RuleKAN's structured Stage 2 search.

KAN-to-PySR achieves a median task-level test NRMSE of 0.0448 (lower is better), compared with 0.0168 for RuleKAN, 0.0105 for RuleSISP, and 0.00107 for PowerRuleKAN-Comp (Appendix Table 7). Instead, PySR fitted directly to observed targets, achieves 0.0170. Thus, KAN-to-PySR does not match our structured methods under the tested settings (the best being PowerRuleKAN-Comp).

Moreover, it is true that PySR supports branching operators, which could help recover fuzzy if–then–else structures. We therefore ran additional PySR searches (using observed targets, not KAN predictions) across all seven fuzzy-rule tasks (three seeds each), adding two operators: a hard if/else operator that selects one branch at a threshold, and a soft-blend operator representing (1−g)u+gv(1-g)u+gv. These additions (slightly) improve predictive accuracy: compared with the original PySR baseline on the same tasks, median task-level test NRMSE decreases from 0.1540 to 0.0976 (lower is better), with improvements on 6/7 tasks. However, this does not translate into better structural recovery. PySR still achieves mean task-median expanded-rule F1 of 0 and 0/21 exact recoveries, versus 0.446 and 8/21 for RuleSISP (Figure 4). Two PySR runs achieve partial recovery, but neither recovers the complete generating structure.

Although PySR frequently selects the new operators, its expressions generally do not reproduce the intended gate-and-branch structure. This further suggests that improving numerical fit does not necessarily recover the generating equation's structure.

We’ll add these considerations and results to the paper. Thank you.

Q2: Contribution of Stage 1 KAN guidance

Yes. Stage 1 KAN guidance restricts symbolic search to candidate variable supports (sets of inputs appearing together in a product term), rather than searching all allowed combinations, as RuleSISP does. This reduces the search space (Appendix C), but missing a required support can prevent recovery of the correct expression. RuleKAN/RuleSISP comparisons show that the effect on predictive accuracy depends on the method variant (Appendix Table 7).

To measure how well KAN guidance identifies the required supports, we evaluated support precision and recall across all 30 analytical tasks (three seeds each). Precision is the proportion of proposed supports matching the generating equation; recall is the proportion of required supports identified. Stage 1 achieves mean task-median precision/recall of 0.440/0.639. After bounded support expansion, which adds candidates derived from KAN evidence without using the generating equations (Section 2 and Appendix F), these scores increase to 0.481/0.983. Thus, Stage 1 identifies many required supports, while the expanded candidate bank covers nearly all of them. The latter measures candidate coverage, not Stage 1 detection alone.

For this rebuttal, to assess whether KAN is preferable to another support detector, we also compared it with multilayer perceptron neural interaction detection (MLP-NID) [2]. Replacing only the source of candidate supports in PowerRuleKAN-Comp, KAN guidance achieves lower symbolic test NRMSE on 17/30 tasks, versus 13/30 for MLP-NID, and median task-level NRMSE of 0.001071 versus 0.001664.

These results support KAN guidance as a useful, but not universally superior, choice for Stage 1. The no-free-lunch theorems for optimization and search [3] establish that no algorithm outperforms all others across all possible problems under their assumptions. Accordingly, different Stage 1 approaches may be better suited to different symbolic regression settings. Our results illustrate this trade-off and motivate considering alternative support detectors rather than assuming a single optimal choice.

We’ll add these considerations and results to the paper. Thank you.

References

  1. Cranmer (2023). Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl. https://arxiv.org/abs/2305.01582
  2. Tsang et al. (ICLR 2018). Detecting Statistical Interactions from Neural Network Weights. https://arxiv.org/abs/1705.04977
  3. Wolpert and Macready (1997). No Free Lunch Theorems for Optimization. https://doi.org/10.1109/4235.585893

Thank you for the work that went into this rebuttal. The additional experiments answer my questions and I maintain my positive assessment of the paper, which I recommend for acceptance.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

As we approach the end of the rebuttal phase, with all three reviewers now supporting acceptance, we would like to briefly summarize the main findings from the additional experiments conducted in response to their concerns:

  1. Generalization beyond our analytical tasks. Evaluations on 100 AI Feynman equations and 24 SRBench datasets confirm that our methods remain competitive with PySR on established benchmarks beyond those originally considered in our paper.

  2. Structured search versus conventional symbolic regression. Applying PySR directly to KAN predictions does not improve over our structured approach. Adding branching operators to PySR slightly improves predictive accuracy on fuzzy-rule tasks, but still yields no exact structural recoveries. This further supports our central finding that numerical accuracy and recovery of the generating symbolic structure are distinct objectives.

  3. Role of KAN guidance. Support precision/recall measurements and comparisons with MLP-NID confirm that KAN guidance is effective, although its advantage over MLP-NID is task-dependent. More importantly, the scalability experiments show that restricting symbolic search to learned supports enables efficient search on a 50-input problem where unrestricted enumeration times out. Thus, the value of KAN guidance lies not only in support identification accuracy, but also in enabling efficient structured symbolic search.

  4. Symbolic expressivity and adaptability. We clarified how the search space can be expanded through configurable operators and validation-guided mechanisms.

We thank all reviewers for their constructive feedback and unanimous support for the paper's acceptance.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.