{"id":"ed6ccf66-4289-4f5f-845a-9d2009521262","arxiv_id":"2508.18173","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GKAN-ODE, a KAN-based symbolic regression model with a dedicated evaluation benchmark, is claimed to recover exact governing equations of graph dynamical systems and beat baselines by up to two orders of magnitude in out-of-distribution trajectory error.","lead":"The paper asks whether AI models can recover the actual equations that drive systems of connected interacting objects, not just imitate their behavior. It introduces a new model, GKAN-ODE, plus a benchmark, and claims the model recovers true equations with big error gains on networks never seen in training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact-recovery claim may conflate library coverage with discovery; need evidence on equations outside the Spline-Wise library.","rationale":"The reader's weakest assumption correctly identifies the central soft spot: the symmetry between the synthetic equation generators and the Spline-Wise symbolic decoder. This is load-bearing because the paper's strongest claim is exact ground-truth recovery, and recovery trivially succeeds when the ground truth is drawn from the same function family the library can express. The OOD generalization to unseen graph topologies is a real but narrower claim; it does not establish generalization over equation space. My concrete test is designed to break this symmetry by introducing a rational term absent from a typical polynomial/trigonometric spline library. If the model still recovers the exact augmented equation, the discovery claim is strengthened; if not, the paper should be scoped as library-constrained symbolic regression. The reader's verdict of CONDITIONAL remains appropriate, since the issue is addressable with additional experiments and code release. I do not see a basis for REJECT, as the paper's contributions to benchmarking and graph-aware modeling are still useful even if the exact-recovery claim requires qualification.","tokens_in":1510,"tokens_out":4248,"duration_ms":50006,"concrete_test":"Generate a synthetic system whose true equation lies outside the Spline-Wise library, e.g., dx_i/dt = a * x_i / (1 + x_i^2) + b * sum_j A_ij * (x_j - x_i). Run the full GKAN-ODE pipeline with the same hyperparameters. If the decoder does not return the exact symbolic form (or an equivalent closed form) and the trajectory error on OOD graphs is not within the claimed two-orders-of-magnitude margin, then the exact-recovery claim is an artifact of library overlap. The paper should report the library's basis functions and per-system recovery success to settle this.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—GKAN-ODE recovers exact ground-truth equations—rests on an unexamined assumption: the synthetic data-generating families and the Spline-Wise symbolic library share the same functional forms. The paper does not specify the library's basis set, nor does it show recovery for equations that cannot be expressed within it. If every ground-truth equation is generated from the same polynomial/trigonometric or other basis that the splines encode, then 'exact recovery' is a curve-fitting consequence, not a discovery of novel symbolic laws. The out-of-distribution test graphs vary topology but preserve the functional family, so the claimed generalization is over graph structure only, not over equation space. Thus, the headline result is compatible with the model interpolating within a predetermined library rather than discovering generalizable governing equations, which is the paper's stated goal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an evaluation pipeline for symbolic regression in graph dynamical systems, focusing on long-term trajectory stability and out-of-distribution generalization to unseen graph topologies. It introduces GKAN-ODE, a graph-adapted Kolmogorov-Arnold network augmented with multiplicative nodes and a Spline-Wise symbolic regression algorithm. The central claim, stated in the abstract, is that GKAN-ODE recovers exact ground-truth equations and achieves trajectory errors up to two orders of magnitude lower than baseline methods on out-of-distribution test graphs. The supplied text contains the abstract, introduction, and references to the proposed methods, but the sections describing the model, the Spline-Wise algorithm, and the experimental results are not provided in the visible manuscript.","tokens_in":1656,"tokens_out":5136,"duration_ms":64469,"significance":"If the claims hold, the work would be a valuable contribution to scientific discovery with graph-structured data: it offers a benchmark focused on generalization and stability, and a neural-symbolic approach that could extract closed-form dynamics from networked systems. The emphasis on out-of-distribution topology generalization is particularly relevant. However, the supplied text provides only a summary-level claim; the actual evidence, methodological details, and experimental results are absent from the visible portion, so the significance cannot be assessed beyond this potential.","major_comments":[{"comment":"The categorical claim that GKAN-ODE 'recover[s] exact ground-truth equations' and achieves 'trajectory errors up to two orders of magnitude lower' is not supported by any quantitative evidence in the supplied text. No per-system results, error bars, or statistical tests are reported. Since this is the paper's central contribution, the experimental section must be included and must provide these details, or the claim should be moderrated.","section":"Abstract / §1"},{"comment":"The Spline-Wise algorithm's function library is not specified. If the synthetic ground-truth equations are generated from the same functional family that the spline library can represent, then 'exact recovery' is a consequence of library coverage, not symbolic discovery. The paper must either show recovery for governing equations outside the spline library (e.g., rational, exponential, or non-smooth terms) or provide an expressiveness analysis that rules out this circularity. Without this, the generalization claim is limited to graph topology, not equation space.","section":"Spline-Wise symbolic regression (described only in abstract)"},{"comment":"The authors propose both the benchmark criteria and the model that wins under them, which creates a risk of benchmarking the model to match the proposed criteria. To strengthen the evaluation, the paper should include strong baselines tuned under the same pipeline, compare against established symbolic regression benchmarks, or show ablations that the pipeline does not implicitly favor GKAN-ODE's design (e.g., the multiplicative node connectivity).","section":"Evaluation pipeline (§1 and abstract)"},{"comment":"The supplied manuscript text truncates after §1, omitting the formal model definition, the Spline-Wise algorithm, the experimental setup, and the results tables/figures. Without these sections, I cannot verify the correctness of the central derivation or the empirical claims. A complete version of the manuscript is required for a thorough review.","section":"§2–§5 (not present in supplied text)"}],"minor_comments":[{"comment":"The phrase 'hyperparameter-free multiplicative nodes' is ambiguous: the connectivity of these nodes (which input pairs are multiplied) is an architectural choice that behaves as a hyperparameter. Please clarify what 'hyperparameter-free' means in this context.","section":"Abstract"},{"comment":"The introduction motivates symbolic discovery but does not position the work relative to existing symbolic regression benchmarks (e.g., AI Feynman, SRBench). Adding a brief comparison would help establish the novelty and the need for a new benchmark.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The provided manuscript text is heavily truncated; if this is an artifact of the submission format, the full PDF should be supplied. The central scientific concern—whether exact recovery is merely library coverage—is not addressed in the visible text and must be handled explicitly. The paper's TMLR acceptance status is noted, but this report is based solely on the supplied material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine step forward for symbolic regression on graph dynamics. It introduces a comparative benchmark with an out-of-distribution-topology criterion, exactly what the field needs, and a KAN-ODE model (GKAN-ODE) with multiplicative nodes and a spline-based decoder that reportedly recovers closed-form equations. If the results hold, this is a practical contribution, not a paradigm shift, but a real one.\n\nThe main claim, however, is too categorical for the evidence available. The abstract says \"recover exact ground-truth equations\" and \"up to two orders of magnitude lower\" errors, but the visible text gives no per-system breakdown, error bars, or statistical tests. The paper appears in TMLR, so presumably peer review caught the worst, but the claim should be backed by open code and a detailed recovery table.\n\nThe stress-test concern about the spline library is on point. If the synthetic ground-truth equations are drawn from the same functional families the splines can express, then \"exact recovery\" measures interpolation within a fixed library, not discovery. The paper needs to show recovery on equations outside the library (e.g., exponential or non-smooth terms) and to report the library's basis set. Without that, the generalization claim is over graph structure only, not over equation space. That doesn't invalidate the benchmark, but it changes what the paper has proven.\n\nI also note the same authors propose the benchmark and the model that wins it. That is a familiar pattern and not fatal, but a third-party baseline or a wider set of baselines would strengthen the evaluation.\n\nOverall, the paper is clearly thought through, the method is sensible, and the evaluation pipeline is a useful standard. The soft spots are in the strength of the headline claims, not in the core idea. I would bring this to a reading group, cite the benchmark when discussing OOD generalization in symbolic regression, and if I were refereeing an earlier version, I would send it out for review—with the specific request that the authors address the library-coverage issue.","headline":"A useful benchmark and KAN-ODE for graph symbolic regression, but the exact-recovery claim needs evidence that the spline library isn't just mirroring the data generators.","tokens_in":2180,"tokens_out":2279,"would_cite":true,"duration_ms":25924,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph-aware neural model recovers exact governing equations for networked dynamical systems and generalizes to unseen topologies.","keywords":["graph dynamical systems","governing equation discovery","symbolic regression","Kolmogorov-Arnold networks","neural ODEs","out-of-distribution generalization","network topology","scientific machine learning"],"falsifier":"Train GKAN-ODE on a graph dynamical system whose true governing equation contains a function outside the spline library (for example, a trigonometric term if the library is polynomial-only) and check whether Spline-Wise symbolic regression still returns the exact equation; if not, the exact-recovery claim is limited to the library's expressive range. A second experiment: train on small random graphs and test on a much larger or denser graph; if trajectory error degrades sharply, the out-of-distribution topology generalization claim is bounded.","tokens_in":1379,"feed_emoji":"🕸️","tokens_out":3429,"duration_ms":42608,"temperature":0.7,"pith_summary":"This paper sets out to show that neural symbolic regression can do more than fit graph dynamics: it can discover the actual closed-form equations that govern them, in a form a scientist can read. To make that case believable, the authors build an evaluation pipeline that judges discovered laws not just by one-step fit but by long-term trajectory stability and by generalization to graph topologies never seen in training. They introduce GKAN-ODE, a Kolmogorov-Arnold-network-based neural ODE with hyperparameter-free multiplicative nodes, plus a Spline-Wise symbolic regression algorithm that decodes the trained network into explicit equations. The central claim is that this model recovers exact ground-truth equations on the tested systems and achieves trajectory errors up to two orders of magnitude lower than baselines when transferred to out-of-distribution test graphs.","feed_headline":"Neural model recovers exact equations for networked dynamics","feed_subtitle":"Trained on known graphs, the model predicts unseen topologies with trajectories up to 100x more accurate.","key_machinery":"The central object is the Graph Kolmogorov-Arnold Network-ODE (GKAN-ODE): a KAN-based neural ODE whose univariate spline activations are decoded into explicit formulas by the Spline-Wise symbolic regression algorithm. Its hyperparameter-free multiplicative nodes let it represent product interactions between connected node states, and the decoding pass greedily converts trained splines into closed-form terms, which is what turns a fitted black box into a recoverable scientific law.","core_discovery":"The paper's central claim is that neural-based approaches, and specifically the GKAN-ODE model, recover exact ground-truth governing equations for graph dynamical systems and generalize far better than existing symbolic regression baselines to previously unseen network topologies. GKAN-ODE adapts Kolmogorov-Arnold Networks to graph dynamics, augments them with hyperparameter-free multiplicative nodes so products of node states can be represented, and adds a Spline-Wise symbolic regression algorithm that turns the learned spline activations into closed-form expressions. The evaluation pipeline is itself part of the contribution: it moves beyond fitting metrics and assesses discovered laws by","pith_inferences":["If recovery is exact only for equations within the spline library's expressive family, the method's discovery power is constrained by that library; the paper's own synthetic generators may have drawn ground-truth equations from the same family the splines can express, so library coverage—not open-ended discovery—could be what guarantees exact recovery.","The out-of-distribution generalization claim suggests the network learns local interaction rules that transfer across global topologies; testing on heterophilic or multiscale networks would show whether that transfer holds when local structure changes.","The Spline-Wise decoding idea is not tied to graphs in principle, so the same train-then-decode approach could be adapted to other structured dynamics, such as spatiotemporal PDEs or multi-agent systems.","A useful falsifying experiment would be to take a graph system whose true right-hand side contains a non-spline function, train GKAN-ODE, and see whether the decoded equation still matches; if it fails, the 'exact recovery' claim is about library coverage rather than generalizability to novel functional forms."],"forward_implications":["Scientists can obtain human-readable symbolic equations for networked dynamics directly from trajectory data, rather than a black-box simulator.","Models trained on one family of graph topologies can be applied to new, unseen topologies without retraining, making the discovered laws transferable across network sizes and structures.","The proposed benchmark—long-term trajectory stability plus out-of-distribution topology generalization—can serve as a standard for evaluating any symbolic regression method on graph dynamics.","Neural symbolic regression becomes a practical alternative to sparse regression for graph coupled systems, with the potential to recover exact laws where classical methods struggle.","Exact recovered equations enable downstream analysis—stability, bifurcations, control design—that is impossible with fitted black-box predictors.","The multiplicative-node design removes a per-dataset hyperparameter, making the approach easier to deploy across different dynamical systems.","If the spline library covers common interaction functions, the pipeline can be applied directly to observational data from epidemic, ecological, or infrastructure networks to propose candidate governing laws.","A natural test is to seed a graph dynamical system with a true term outside the spline library and check whether Spline-Wise regression still recovers the exact equation; failure would show that 'exact recovery' is bounded by library coverage, not open-ended discovery."],"supporting_citations":[],"fun_headline_variants":["GKAN-ODE: exact equations for unseen network topologies","Neural networks discover governing equations that generalize","Benchmark: neural models beat symbolic regression on graphs","AI model discovers exact laws for unseen graph topologies","GKAN-ODE beats baselines on unseen networks"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ground-truth dynamics are assumed to be representable by the spline function library that the Spline-Wise decoder can turn into equations, so 'exact recovery' may measure coverage of that library rather than open-ended symbolic discovery.","fun_headline_variants_meta":{"raw":{"variants":["GKAN-ODE: exact equations for unseen network topologies","Neural networks discover governing equations that generalize","Benchmark: neural models beat symbolic regression on graphs","AI model discovers exact laws for unseen graph topologies","GKAN-ODE beats baselines on unseen networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001101,"raw_usage":{"total_tokens":4424,"prompt_tokens":736,"completion_tokens":3688,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":3612}},"tokens_in":480,"tokens_out":3688,"duration_ms":29807,"temperature":1.0,"reasoning_tokens":3612,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:31:39.163136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GKAN-ODE on a graph dynamical system whose true governing equation contains a function outside the spline library (for example, a trigonometric term if the library is polynomial-only) and check whether Spline-Wise symbolic regression still returns the exact equation; if not, the exact-recovery claim is limited to the library's expressive range. A second experiment: train on small random graphs and test on a much larger or denser graph; if trajectory error degrades sharply, the out-of-distribution topology generalization claim is bounded.","supporting_citations":[],"review_version":1}