{"id":"14b9a8ac-b38b-4d7a-b12e-b7f65f5998d3","arxiv_id":"2504.16503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EN4SR couples evolutionary topology search with gradient-based weight tuning and a reusable weight memory, and beats NN-only symbolic regression baselines in reported experiments.","lead":"This paper presents EN4SR, a symbolic regression algorithm that evolves neural network topologies with genetic operators and tunes their weights with gradient descent. It reports more accurate models than neural-only baselines on three physics benchmarks and a quadcopter system identification task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Budget parity is unverified: without measured wall-clock time or actual total backprop iterations, the claim that EN4SR beats N4SR at comparable compute is not established.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue. The paper's stated central contribution includes 'as good as or better than the pure NN-based approach in terms of model accuracy and computational demands' (Section 1), and Section 5.3 attempts to guarantee fairness by setting a maximum of 90,000 backprop iterations. However, the paper never verifies how many backprop iterations EN4SR actually consumes. Algorithm 1's total cost depends on the unspecified size of the offspring set produced by generateSubtopologies and on the number of nondominated solutions fine-tuned each generation. The quality comparison on the three benchmarks is otherwise reasonably supported: 30 runs, medians, Wilcoxon p-values, and an ablation with EN4SR-base. The quadcopter section is illustrative rather than comparative. The separate overclaim about significant superiority over mSNGP-LS on magic and magman is real but secondary; it should be corrected in the text. The budget-parity concern is concrete and testable by instrumenting the released code, so the existing CONDITIONAL verdict is the right one and no verdict change is needed.","tokens_in":14901,"tokens_out":4987,"duration_ms":48632,"concrete_test":"Run the released code (https://github.com/jirkakubalik/EN4SR) on the magic problem with the reported settings and instrument Algorithm 1 to log (a) total Adam update steps per run across all populations and (b) wall-clock time. Log the same for N4SR-ACYE under its 90,000-iteration cap. Repeat for 30 seeds. If EN4SR's median total update steps exceeds 90,000 by more than 5%, or its median wall-clock time exceeds N4SR's by more than 10%, the 'similar computing budget' premise in Section 5.3 fails and the quality comparison is confounded by compute.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 states 'the maximum number of backprop iterations was set to 90000 for a fair comparison with N4SR,' but the paper reports no wall-clock times and no actual total backprop-iteration counts for any method. Algorithm 1's total cost is not fixed by the listed parameters: line 9 (`interPop ← pop.generateSubtopologies(N_n, L1)`) does not specify how many subtopologies are generated per generation, and lines 11–12 fine-tune an unspecified set of nondominated solutions for N_f steps every generation. With POP_SIZE=10, R=3, G=20, N_n=10, N_t=100, N_f=50, even a minimal reading (10 offspring per generation, at most 10 nondominated solutions) yields roughly 99,000 Adam steps for EN4SR, so the '90,000' cap appears to apply to N4SR only. If the offspring count is larger, EN4SR's budget exceeds N4SR's by a wide margin. Therefore the central comparison—EN4SR 'outperforms' N4SR in model quality—is potentially explained by additional compute rather than by the neuro-evolutionary mechanism. The contribution claim that EN4SR is 'as good as or better ... in terms of ... computational demands' (Section 1) is unsupported without runtime or iteration measurements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EN4SR, a neuro-evolutionary symbolic regression method that wraps the N4SR neural-network architecture in an evolutionary search over network topologies, using a weight memory and periodic perturbations to allow short backpropagation runs for each candidate. The method is evaluated on three physics-aware benchmark problems (resistors, magic, magman) against N4SR-ACYE, EQL-divide, and mSNGP-LS, plus a real-world quadcopter system-identification case study. The central claim is that EN4SR matches or exceeds pure NN-based symbolic regression in model accuracy and computational demand, and significantly outperforms GP-based regression on model quality. Table 1 reports median RMSE values and Wilcoxon p-values that support the accuracy advantage on the three benchmarks, but the computational-cost claim is not backed by any runtime or actual-iteration measurement, and the quadcopter section contains no quantitative error metric.","tokens_in":15120,"tokens_out":3505,"duration_ms":35479,"significance":"If the main claims hold, EN4SR is a practically useful extension of NN-based symbolic regression: the evolutionary wrapper appears to improve final model accuracy without requiring full training of each candidate topology, and the memory mechanism is a sensible way to reuse weight information across generations. Strengths of the paper are that the source code is publicly available, the evaluation uses external benchmark problems rather than fitted examples, and the statistical testing via Wilcoxon rank-sum tests is appropriate for the reported medians. However, the significance of the contribution is currently limited by the lack of measured compute budgets, by the absence of any distributional information beyond medians, and by the qualitative nature of the quadcopter demonstration; these issues prevent the reader from verifying the headline claim that EN4SR is better than NN-based approaches 'with a similar computing budget.'","major_comments":[{"comment":"","section":"Section 5.3 and Algorithm 1"},{"comment":"","section":"Table 1"},{"comment":"","section":"Section 5.5"},{"comment":"","section":"Section 6"}],"minor_comments":[{"comment":"","section":"Section 4.1"},{"comment":"","section":"Table 1 caption and Section 5.4"},{"comment":"","section":"Figure 3 and Section 5.4"},{"comment":"","section":"Section 5.3"},{"comment":"","section":"Section 5.5"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' prior N4SR work, both conceptually and as the baseline, and the novelty is mainly the evolutionary wrapper and memory mechanism. This is within scope for GECCO, but the authors should be asked to make the compute-budget comparison explicit, as the current form leaves the main comparative claim open to the alternative explanation that EN4SR simply uses more gradient steps."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe short version: EN4SR is a reasonable incremental extension of the authors' own N4SR work, with a real new idea in the weight memory, and the experiments mostly support the accuracy claim. The soft spot is the compute-parity claim, which is not backed by any runtime or iteration measurement.\n\nWhat's new: the paper couples evolutionary topology search to the N4SR-style gradient-trained network, adds a weight-memory mechanism to seed offspring with learned weights, and alternates stages with perturbations. That's a sensible design for small-data physics-aware symbolic regression. The ablation (EN4SR vs EN4SR-base) gives some evidence that the memory helps, at least on magman. The code is public, and the comparisons use Wilcoxon p-values rather than just point estimates.\n\nWhere it gets shaky: the budget-parity claim. Section 5.3 says the max backprop iterations were capped at 90000 for fairness, but no wall-clock times or actual iteration counts are reported. Algorithm 1 doesn't specify how many offspring are generated per generation, and with the listed parameters (R=3, G=20, Nn=10, Nt=100, Nf=50) a plausible reading puts EN4SR's total Adam steps well north of 90000. If that's right, the 'comparable compute' claim is not established, and the quality advantage could partly be a compute advantage. The authors need to report total per-run iterations or wall-clock, and clarify what the 90000 cap applies to.\n\nThere's also an overclaim in the contributions bullet: it says EN4SR significantly outperforms the GP baseline, but the table shows non-significant p-values on magic (0.21) and magman (0.17). The body text is more careful ('performs equally well'), so the abstract/contributions are just sloppy. Minor: no error bars or distributional measures (only medians), and the quadcopter demo is qualitative—no RMSE or comparison.\n\nOverall, the central accuracy claim is probably fine for what it is, but the paper oversells the compute and GP-significance points. A referee should ask for those numbers before accepting.\n\nWho's this for? People working in symbolic regression with physics priors, especially if they build on N4SR. It's a solid contribution that deserves a serious look, but the revisions should be required, not cosmetic.\n\nRecommendation: send to peer review, but with a request for measured compute and corrected significance claims.","headline":"EN4SR is a solid incremental extension of N4SR with a useful weight-memory trick, but the compute-parity claim needs real measurements before the efficiency contribution can be believed.","tokens_in":15697,"tokens_out":5139,"would_cite":true,"duration_ms":44155,"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":"EN4SR combines evolutionary topology search with gradient tuning to beat pure NN symbolic regression on physics-aware benchmarks.","keywords":["Symbolic regression","Neuroevolution","Physics-aware modeling","Multi-objective optimization","Neural topology search","Weight memory","Gradient-based tuning","Dynamic system identification"],"falsifier":"Log the total wall-clock time and the true number of gradient evaluations per run for EN4SR and N4SR-ACYE on the three benchmark problems, instead of relying only on the 90,000-iteration cap; if EN4SR consistently consumes substantially more compute per run, the claim of comparable computational demand is not supported. A secondary check would compare EN4SR against EN4SR-base across many more runs to see whether the weight memory's contribution, which is significant only on magman in the paper, appears on the other problems as well.","tokens_in":14641,"feed_emoji":"🧬","tokens_out":9963,"duration_ms":90199,"temperature":0.7,"pith_summary":"This paper proposes EN4SR, a symbolic regression method that evolves neural-network topologies with genetic operators while fitting the network weights with short runs of gradient descent. The intended payoff is a physics-aware data-driven modeling tool that avoids the premature convergence to suboptimal structures that limits pure neural-network symbolic regression, without paying the full cost of training every candidate topology. On three benchmark problems from prior work—resistor networks, tire-road friction, and magnetic force—the paper reports that EN4SR outperforms the neural-network baseline N4SR-ACYE in interpolation-plus-extrapolation accuracy under a fixed backpropagation budget, and matches or beats a GP-based method. A quadcopter identification example shows the method recovering a simple linear velocity model from real flight data.","feed_headline":"Evolving network shapes beats pure NN symbolic regression","feed_subtitle":"A hybrid that evolves neural topologies and tunes weights by backprop beats pure NN models under a fixed budget.","key_machinery":"The load-bearing components are the master-topology/subtopology scheme, the weight memory, and the staged perturbation. The master topology is a fixed heterogeneous feedforward network whose units are elementary functions and whose layers include copy units connected by skip connections; a subtopology is any sparse subnetwork derived by enabling and disabling units and links. The weight memory stores the z-node weight vectors of recent well-performing, non-dominated subtopologies, organized per unit, and crossover and mutation draw from that memory with probability $p_h$, so freshly created topologies start from parameter values that have already worked. Each stage begins by perturbing every population member, enabling all units and reinitializing their weights from memory or at random, which the paper argues prevents irreversible stagnation. Candidate subtopologies receive few backprop steps ($N_n=10$), survivors are tuned more ($N_t=100$), and non-dominated solutions are fine-tuned ($N_f=50$) with progressively richer loss functions that combine training, singularity, constraint, and sparsity terms.","core_discovery":"EN4SR's central claim is that a neuro-evolutionary hybrid, evolutionary search over network topologies combined with gradient-based weight tuning and supported by a weight memory plus periodic population perturbation, can discover physics-aware symbolic models of higher quality than a pure neural-network symbolic regressor under the same backpropagation budget. The paper defines higher quality as lower root-mean-square error on held-out data covering both interpolation and extrapolation regions, at comparable or slightly higher model complexity. This result is reported on all three proof-of-concept problems with statistical significance against N4SR-ACYE, and the method is further demonstrated on a quadcopter dynamics identification task where most runs converge to the same linear model.","pith_inferences":["Beyond the paper, the weight-memory design could be borrowed by any evolutionary-neural system: storing parameter vectors of useful subgraphs and re-injecting them after crossover or mutation is a general way to cut retraining cost, not a trick specific to symbolic regression.","Beyond the paper, the final-model selection rule is a bottleneck the paper itself flags: the algorithm returns the least-complex model below median performance, so the best non-dominated model in the population may be systematically under-reported, and a selection rule that searches the whole front more carefully could improve reported accuracy without changing evolution.","Beyond the paper, the quadcopter result suggests a testable prediction: on outdoor flights with wind, the same method should recover the same linear model plus an affine offset estimating the wind-velocity component; if it does not, the zero-offset outcome simply reflects the indoor dataset rather than a modeling principle."],"forward_implications":["Under the same 90,000-iteration backprop budget, EN4SR reports lower median $\\mathrm{RMSE}_{\\mathrm{int+ext}}$ than N4SR-ACYE on all three benchmark problems, with the differences reaching statistical significance.","EN4SR matches or beats the GP-based mSNGP-LS on magic and magman and significantly outperforms it on resistors.","The memory component's contribution is not uniform: EN4SR-base without memory is statistically indistinguishable from EN4SR on resistors and magic but significantly worse on magman ($p=0.015$).","On the quadcopter task, 24 of 30 runs converge to the same linear model $v_x(k+1)=0.985\\,v_x(k)+0.473\\,\\theta(k)$, which tracks an unseen square-trajectory test set in recursive simulation.","EN4SR's models are slightly more complex than N4SR-ACYE's on the benchmarks, trading a small complexity increase for accuracy."],"supporting_citations":[{"why":"Supplies the master-topology architecture, the three-stage loss composition, and the N4SR-ACYE baseline that EN4SR is evaluated against.","marker":"[23]"},{"why":"Defines the synthetic constraint-sample scheme used to enforce prior-knowledge constraints in multi-objective selection.","marker":"[22]"},{"why":"Provides the sparsity-promoting regularization used to drive learned networks toward concise expressions.","marker":"[15]"},{"why":"Adam is the gradient-based optimizer used for the short backprop runs on subtopologies.","marker":"[16]"},{"why":"Introduces the resistors benchmark and the symmetry, diagonal, and upper-bound constraints used in the first test problem.","marker":"[6]"},{"why":"Provides the empirical magnetic-force model used to generate extrapolation data for the magman problem.","marker":"[14]"},{"why":"Supplies the 'magic formula' tire-road model and its parameter values used as the reference for the magic problem.","marker":"[38]"}],"fun_headline_variants":["Hybrid evolution beats pure NN for symbolic regression","Evolving topologies outperform pure NN in symbolic regression","Memory-guided evolution beats pure NN symbolic regression","Physics-aware regression improved by neuroevolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The budget-parity premise: capping both methods at the same number of backprop iterations makes their total computational costs comparable, even though EN4SR's evolutionary operations add overhead that is never measured in wall-clock time.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid evolution beats pure NN for symbolic regression","Evolving topologies outperform pure NN in symbolic regression","Memory-guided evolution beats pure NN symbolic regression","Physics-aware regression improved by neuroevolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001323,"raw_usage":{"total_tokens":5355,"prompt_tokens":887,"completion_tokens":4468,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":4411}},"tokens_in":503,"tokens_out":4468,"duration_ms":29389,"temperature":1.0,"reasoning_tokens":4411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:01:45.466144+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Log the total wall-clock time and the true number of gradient evaluations per run for EN4SR and N4SR-ACYE on the three benchmark problems, instead of relying only on the 90,000-iteration cap; if EN4SR consistently consumes substantially more compute per run, the claim of comparable computational demand is not supported. A secondary check would compare EN4SR against EN4SR-base across many more runs to see whether the weight memory's contribution, which is significant only on magman in the paper, appears on the other problems as well.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adam is the gradient-based optimizer used for the short backprop runs on subtopologies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the empirical magnetic-force model used to generate extrapolation data for the magman problem."},{"cited_title":"Verdier, Robert Babuška, Barys Shyrokau, and Manuel Mazo","cited_arxiv_id":null,"evidence_quote":"Supplies the 'magic formula' tire-road model and its parameter values used as the reference for the magic problem."}],"review_version":1}