- 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
-
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?
-
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?
-
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
For clarity, "" is the expanded sum-of-products form of the nested fuzzy tree. Factoring out from the first two terms reveals the nested gating hierarchy ( followed by ).
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.
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) can identify support {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 d inputs and fixed factor order q, RuleSISP's Stage 2 search space grows as O(dq), 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=3) and 8.71 quintillion at 50 inputs (q=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).
Its median test NRMSE is 9.13×10−8, 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:
3.02303tanh(x4exp(x0))+exp(−x1−cos(x2))+exp(x3+exp(x0))1.
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