- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Logic-KAN, which learns executable Boolean programs from labeled binary samples. The model uses differentiable parameters to select positive literals, negative literals, and omission states, constructs shared conjunctive clauses, and then combines them with an OR/XOR head. During training, gradient updates alternate with discrete structural edits; after training, it further performs program simplification and data-based full candidate selection. The final program is checked by an independent executor and a SAT solver to evaluate compilation consistency and target-function equivalence, respectively.
Critical strengths
- The exact recovery results on DNF are convincing. Across two batches of external comparisons, Logic-KAN recovers 95/100 targets in total, whereas RRL, despite exceeding 99.9% test accuracy, still achieves 0/100 exact recovery. This shows that the metric the paper focuses on does reveal differences that ordinary accuracy struggles to capture.
- The local edit scoring has a clear motivation and experimental support. Fixing the outer OR masks the correction of some subprograms, and the paper provides a clear analysis of this issue. In the paired ablation, replacing local scoring with global scoring reduces exact recovery from 47/50 to 20/50, supporting the role of this design.
- The verification pipeline is clearly differentiated. The authors separately check the consistency between the continuous and discrete models, between the discrete model and the derived program, and between the derived program and the target function, without equating compilation correctness with successful rule recovery. The blind test on new targets after development, which reaches 147/150, further strengthens the credibility of the results.
Critical weaknesses
- Some of the successful results rely mainly on specialized post-training candidate construction, and the contribution of the core learning process needs to be further isolated. In Table 11, after removing the affine-XOR, signed-count, and positive-closure candidates, exact recovery on Parity16, Majority7, and Conjunction drops from 5/5, 5/5, and 3/5 to 0/5, respectively. These candidates incorporate direct construction algorithms tailored to the corresponding function families. Using them is reasonable, but a baseline with "candidate generation and selection only" is currently missing, making it impossible to judge how much additional benefit the preceding gradient learning and structural editing contribute on these tasks.
- The main Mixed comparison is insufficient to verify OR/XOR compositional capability. Appendix A.5 shows that the functions produced by this generator can be simplified to seven-clause DNF. Therefore, full recovery on Mixed does not require learning the original XOR structure, and BRCG also achieves 100/100. The added Random XOR blind test provides some supplement, but the external method comparison still lacks tasks that genuinely exhibit the advantages of mixed structure.
- The claim about program compactness lacks direct evidence. The abstract claims that Logic-KAN produces smaller programs than existing methods on both task types, but the main text and appendix mainly report structural statistics of this work and do not provide a comparison of program size across methods on the same tasks. Recovering the teacher's clause set is not equivalent to proving that the program is smaller. This gap is especially important because the growth-only control reaches 50/50, higher than the full method's 47/50; if the full method's advantage lies in size, the corresponding trade-off should be shown directly.
- The comparison does not yet fully rule out the effects of search budget and post-processing. In addition to gradient training, Logic-KAN uses multiple rounds of structural search, post-training simplification, and full candidate selection, whereas the computational budgets of the methods are not matched. The paper also lacks a purely discrete search control under the same grammar and candidate budget. It is therefore still difficult to judge how much of the advantage comes from alternating gradient and editing, and how much comes from additional search and post-processing.
Questions for authors
- If neural training is skipped and only the affine-XOR, positive-closure, and signed-count candidate generators are run with the same selection criteria, what are the results on the classical tasks? Could the recovery rates for these tasks also be reported at the checkpoint, after simplification, and after candidate selection?
- What are the differences between the full method and growth-only in terms of program size, recovery rate, and runtime? Under the same grammar and search budget, can purely discrete editing achieve comparable results?
- Could a comparison of program size across methods on the same tasks be added, along with an explanation of how programs are counted uniformly across different representations? I would also like to see external baseline results on tasks that genuinely require XOR composition and cannot be easily simplified into small DNF.