{"id":"ffd2356a-f330-4b63-b03e-82b7d5d40731","arxiv_id":"2608.10235","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A matched-integrator, multi-seed evaluation finds Hamiltonian Neural Networks reduce long-horizon energy drift by 42x and trajectory MSE by 15.8x on the pendulum and retain a smaller but clear advantage on 3D Kepler.","lead":"This paper runs a tightly controlled comparison between two kinds of neural networks that learn physics: a Hamiltonian Neural Network, which learns an energy-like function, and an ordinary network that learns motion directly. Under identical training and the same integrator, the Hamiltonian network shows far smaller energy drift and trajectory error on a pendulum and on the 3D Kepler two-body problem, but the advantage is scoped to simple integrable systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The matched-integrator control does not fully separate the Hamiltonian prior from the scalar-output/autograd implementation; an ablation is needed to attribute the 42x/15.8x ratios.","rationale":"The reader's weakest assumption correctly identifies the load-bearing point: the matched-integrator protocol compares a direct vector-field network with a scalar-output network whose vector field is a symplectic gradient. These differ not only in the Hamiltonian prior but also in output dimensionality and in the use of autograd, so the observed advantage is not unambiguously attributable to the prior alone. I find no critical mathematical error; the paper's own limitations (Section 7) are honest, and the bounded-versus-growing energy curves and the Kepler transfer are plausibly explained by the Hamiltonian parameterization. However, the absent control (a scalar-output non-Hamiltonian baseline, or a divergence-regularized direct baseline) is the cost of accepting the strong quantitative claim. The proposed ablation would settle whether the 42x/15.8x numbers are due to the symplectic-gradient constraint or to the scalar-potential implementation. Secondary concerns such as the wide seed variance and the horizon-dependence of the ratios are acknowledged in the paper and do not threaten the qualitative finding. Therefore the verdict remains CONDITIONAL, pending the ablation and the release of code/data, exactly as the reader concluded.","tokens_in":12825,"tokens_out":14336,"duration_ms":145431,"concrete_test":"Train a third model with the same architecture as Model B (scalar output, autograd) but with the vector field defined as the Euclidean gradient (∂q φ, ∂p φ) instead of the symplectic gradient (∂p φ, −∂q φ). Use the same data, seeds, optimizer, and RK4 rollout as the main experiment, and report the T=100 mean energy drift and trajectory MSE over five seeds. If this 'gradient-field' control behaves like Model A (large growing drift), the Hamiltonian/symplectic structure is the cause; if it behaves like Model B (bounded low drift), the scalar-output/autograd pipeline is the cause. Alternatively, add a divergence-penalty term to Model A's loss and test whether it closes the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Under the protocol of Sections 3.3 and 3.4, Models A and B share data, derivative targets, optimizer, RNG seed schedule, and the RK4 inference integrator, and the shared hidden layers are initialized identically given the seed. However, Model B differs from Model A in two coupled ways: it outputs a single scalar (65 fewer parameters, Section 3.2) and its vector field is obtained by autograd through Eq. (4), imposing the symplectic-gradient form (∂p Hθ, −∂q Hθ). The paper attributes all improvement to the Hamiltonian architectural prior, but this prior is implemented via the scalar-output/autograd pipeline. If that pipeline independently changes the optimization landscape or the function class of the vector field (e.g., producing a smoother or more regular field), then the 42x energy-drift and 15.8x trajectory-MSE ratios would not be a clean isolation of the Hamiltonian mechanism. The control is strong but incomplete: no baseline separates the symplectic-gradient constraint from the scalar-potential representation, so the attribution of the headline ratios to the Hamiltonian prior remains conditional. This is the weakest link in the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper evaluates Hamiltonian Neural Networks against a standard feedforward network under a matched-integrator protocol: identical RK4-generated training data, identical central-difference derivative targets, the same optimizer settings and seed schedule, and the same RK4 integrator at inference, with five independent training seeds. On the nonlinear pendulum the HNN is reported to reduce mean maximum energy drift by 42x and mean trajectory MSE by 15.8x at T=100, with bounded rather than growing drift and lower seed-to-seed variance; an energy-stratified analysis shows the gap widening for more nonlinear orbits. A Störmer–Verlet rollout of the learned HNN is analyzed as a diagnostic and does not improve on HNN+RK4. The protocol is then transferred to the three-dimensional Kepler two-body problem, where the HNN shows smaller but consistent advantages in trajectory MSE, energy drift, and angular-momentum drift. The paper also reports computational costs and discusses limitations and future work.","tokens_in":13072,"tokens_out":7163,"duration_ms":69512,"significance":"If the results hold, the paper provides a carefully controlled confirmation of the HNN architectural prior on conservative dynamics. Its strengths include the matched inference integrator, identical supervision targets, multi-seed statistics, the bounded-versus-growing energy-drift diagnostic, the energy-stratified analysis, the honest negative Verlet result, the computational-cost comparison, and a transfer experiment in six-dimensional phase space with an additional conserved quantity. The reported ratios, however, rest on an attribution that is not fully isolated: the HNN differs from the baseline both in its scalar-output/autograd implementation and in the symplectic-gradient structure, and the headline pendulum ratios are sample-mean estimates with large per-seed variance. These issues are fixable with additional control experiments and uncertainty quantification, and they do not undermine the qualitative bounded-drift phenomenon.","major_comments":[{"comment":"The matched-integrator protocol does not fully isolate the Hamiltonian prior from the scalar-output/autograd implementation. Model B differs from Model A in two coupled ways: it has a single scalar output (65 fewer parameters, §3.2) and its vector field is obtained by automatic differentiation of that scalar through Eq. (4), which imposes the symplectic-gradient form. The claim in §7 that 'the only intentional difference is that Model B's vector field is the symplectic gradient of a learned scalar rather than a direct prediction' conflates the architectural prior with its implementation. Because the scalar-output representation and the autograd pipeline could, on their own, change the optimization landscape or the smoothness of the learned field, the 42× and 15.8× ratios are not yet cleanly attributable to Hamiltonian structure. I request an additional control: train a scalar-output network of the same width and depth whose vector field is (∂_p s_θ, ∂_q s_θ) (or another non-Hamiltonian gradient sign) so that the scalar-output/autograd cost is identical but the symplectic structure is absent; if this baseline behaves like Model A, the ratios can be attributed to the Hamiltonian prior, and if it behaves like Model B, the paper's central interpretation needs to be revised.","section":"§3.3–3.4, §7"},{"comment":"The headline ratios of 15.8× and 42× are ratios of sample means over only five seeds, and Model A's seed-to-seed variance is very large at T=100 (σ/μ > 100% for trajectory MSE and ≈85% for energy drift). No uncertainty interval, paired per-seed ratio, or significance test is reported for these ratios, so the point estimates may be much less stable than the abstract suggests. Please report per-seed ratios (or paired differences on a log scale) and bootstrap confidence intervals for the ratios at each horizon, and state whether the bounded-versus-growing claim is assessed on paired or aggregate statistics. This is important because the quantitative ratios are the paper's central empirical contribution.","section":"§4.3, Tables 2 and 3"}],"minor_comments":[{"comment":"The abstract and table call Model B 'parameter-matched' although it has 65 fewer parameters than Model A; since 65/8642 is small this is a minor wording issue, but 'near-parameter-matched' or an explicit statement of the difference would be more precise.","section":"Abstract, Table 1"},{"comment":"In the low-energy tercile at T=10, Model B's trajectory MSE (4.56±2.65)×10−6 is worse than Model A's (1.24±1.16)×10−6, which is opposite to the energy-drift advantage; the text's claim that the architectural advantage widens with nonlinearity should acknowledge this exception.","section":"Table 4"},{"comment":"The HNN energy-drift entries for T=10 through T=100 are identical to two significant figures ((2.16±0.90)×10−4); please clarify whether this is rounding of a saturated value and report additional significant digits so readers can verify the saturation claim.","section":"Table 3"},{"comment":"The description of initial conditions as 'q0 uniform on spheres of radius r0 ∈ [0.9,1.4]' is ambiguous: specify whether the radius is drawn uniformly and the direction uniformly on the sphere, or whether a different sampling over the shell is used.","section":"§6"},{"comment":"The phrase 'same optimiser state schedule' is imprecise because Adam state tensors have different shapes for the two models; saying 'same optimizer hyperparameters and update schedule' would be accurate.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of cs.LG and the experiments are described with unusual care. My main concern is the missing ablation separating the Hamiltonian prior from the scalar-output/autograd implementation, along with the lack of uncertainty quantification on the headline ratios; both are addressable within the manuscript's scope. If the authors add the control experiment and tighten the statistics, I would be willing to support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a carefully controlled benchmark study of Hamiltonian neural networks on pendulum and Kepler problems, and the control protocol is the real contribution. Second, the headline 42x energy-drift and 15.8x trajectory-MSE numbers are credible, but the attribution to the Hamiltonian prior is slightly less clean than the prose suggests because the HNN implementation also changes the output head and uses autograd.\n\nWhat the paper does well: both models share data, derivative targets (central difference), optimizer, seed schedule, and the same RK4 integrator at inference. Five seeds, mean±std. The HNN's energy drift saturates and stays bounded; the standard NN's grows. The energy-stratified analysis shows the advantage widens in more nonlinear regimes. They also report a sample-efficiency sweep with a low-data crossover, wall-clock costs, a negative result for an explicit Verlet rollout (correctly noting the learned Hamiltonian need not be separable), and a 6D Kepler transfer where the advantage persists at smaller magnitude and angular-momentum drift is also lower. They explicitly credit Greydanus et al. for the original pendulum finding and are appropriately modest about scope. The limitations section is honest.\n\nSoft spots: the matched-integrator protocol does not fully isolate the Hamiltonian prior. The HNN outputs a scalar and obtains its vector field through autograd, which changes the function class and optimization landscape independently of the symplectic sign structure. The authors call autograd a consequence of the architecture, which is fair, but an ablation (e.g., a scalar-output network whose gradient is used with a non-Hamiltonian sign convention) would make the causal claim airtight. Without that, the 42x is an upper bound on the effect of the full Hamiltonian parameterization, not a clean estimate of the prior alone. Minor issues: no code/data released yet; the pendulum ratio has no uncertainty (Kepler ratios do); the trajectory-MSE ratio dips at T=10, which they explain as noise but it reminds you the ratios are noisy. None of these are fatal.\n\nWho should read it: anyone comparing physics-informed architectures or building on HNNs. It deserves a proper peer review; I'd ask for the ablation and code/data. It's a useful benchmark paper, not a breakthrough, and it doesn't pretend otherwise.\n\nRecommendation: send it out; require the ablation and artifacts before acceptance.","headline":"A genuinely careful matched-integrator benchmark that sharpens the known HNN advantage; the control is strong, but an ablation is needed before calling the 42x ratio purely the Hamiltonian prior.","tokens_in":13596,"tokens_out":4147,"would_cite":true,"duration_ms":41435,"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":"Hamiltonian prior cuts learned-model energy drift 42-fold","keywords":["Hamiltonian neural networks","energy drift","matched-integrator evaluation","pendulum dynamics","Kepler two-body problem","symplectic integration","physics-informed machine learning","long-horizon prediction"],"falsifier":"Train a third architecture under the identical protocol—a scalar-output network whose two-dimensional vector field is produced by multiplying the gradient by a fixed, non-Hamiltonian matrix—and compare long-horizon energy drift; if its drift also saturates near $2\\times10^{-4}$, the reported $42\\times$/$15.8\\times$ ratios would be an effect of scalar parameterization rather than of Hamilton's equations.","tokens_in":12651,"feed_emoji":"🪐","tokens_out":8631,"duration_ms":73521,"temperature":0.7,"pith_summary":"This paper tries to establish that the Hamiltonian architectural prior—learning a scalar energy function and deriving the dynamics from its gradient—is what improves long-horizon prediction, rather than any side benefit of the training pipeline. Under a matched protocol where a Hamiltonian network and a standard feedforward network share the same RK4-generated trajectories, the same central-difference derivative targets, the same optimizer settings, and the same RK4 integrator at inference, the Hamiltonian network reduces mean energy drift by 42 times and mean trajectory error by 15.8 times at $T=100$, roughly 16 pendulum periods, over five independent seeds. Its energy drift stays bounded while the baseline's grows, and the advantage widens on more nonlinear orbits. If the claim holds, it means the structural prior buys both accuracy and predictability: the Hamiltonian network's long-horizon behavior is not only closer to the true dynamics but far less sensitive to seed. The same protocol transfers to 3D Kepler with smaller but consistent ratios on energy, angular momentum, and trajectory error.","feed_headline":"Hamiltonian prior cuts learned-model energy drift 42-fold","feed_subtitle":"Matched data, optimizer, and integrator isolate the prior: drift stays bounded over 16 pendulum periods.","key_machinery":"The load-bearing object is the learned scalar Hamiltonian $\\hat{H}_\\theta(q,p)$ whose symplectic gradient $(\\partial_p \\hat{H}_\\theta, -\\partial_q \\hat{H}_\\theta)$ supplies the vector field; this is what makes the dynamics Hamiltonian by construction. The matched-integrator protocol—identical RK4-generated data, identical central-difference derivative targets, the same Adam settings and seed schedule, and the same RK4 stepper at inference over five seeds—is the control that lets the paper attribute the measured ratios to that architectural prior. The bounded-versus-growing energy-drift signature and the energy-stratified tercile analysis are the diagnostics showing where the prior acts: the standard network's drift grows with horizon and with orbital nonlinearity, while the Hamiltonian network's stays roughly constant.","core_discovery":"On the paper's own terms, the central discovery is that, once data, supervision, optimization, and inference integration are held fixed, the Hamiltonian parameterization itself accounts for the long-horizon advantage. On the nonlinear pendulum, the learned Hamiltonian's mean maximum energy drift saturates near $2.16\\times 10^{-4}$ by $T=5$ and stays flat through $T=100$, while the matched standard network's drift grows monotonically from $2.75\\times 10^{-4}$ to $9.08\\times 10^{-3}$, a $42\\times$ ratio of means at $T=100$; mean trajectory MSE is $15.8\\times$ smaller. The learned scalar $\\hat{H}_\\theta$ matches the analytical pendulum Hamiltonian to RMSE $7.3\\times 10^{-5}$ on the training domain, and the saturated drift is the same order as that Hamiltonian-approximation error. In 3D Kepler, the same protocol gives $8.0\\pm 1.0\\times$ smaller energy drift, $4.2\\pm 0.5\\times$ smaller angular-momentum drift, and $5.5\\pm 0.8\\times$ smaller trajectory MSE at $T=10$, even though angular momentum is not in the training loss. The paper also reports a negative diagnostic: an explicit Störmer–Verlet rollout of the learned, non-separable Hamiltonian does not beat HNN+RK4, because the symplecticity guarantee of velocity Verlet only applies to separable Hamiltonians.","pith_inferences":["A testable extension follows from the bounded-drift mechanism: if the benefit comes from scalar parameterization rather than Hamilton's equations, then other scalar-energy architectures (spectral or Gaussian-process Hamiltonians) should reproduce the same saturation under the same protocol.","The Kepler angular-momentum result is empirical evidence that rotational symmetry is partially captured without explicit equivariance; an explicit rotationally equivariant Hamiltonian parameterization would be the natural test of whether the residual drift closes further.","The low-data crossover at eight training trajectories is a boundary the paper notes but does not explain; a mechanistic account of when the Hamiltonian prior becomes identifiable would turn this into a practical selection rule.","Chaotic benchmarks named in the paper, such as Hénon–Heiles or the double pendulum, are the sharpest test of whether bounded drift is a property of integrable systems or of the Hamiltonian prior itself."],"forward_implications":["At $T=100$ on the pendulum, the Hamiltonian network's energy drift is bounded near $2\\times10^{-4}$ while the baseline's grows to $9\\times10^{-3}$, so practitioners can trust HNN rollouts over about 16 periods with five-seed-consistent behavior.","The architectural advantage widens with nonlinearity: from low- to high-energy terciles the baseline's drift grows $4.3\\times$ while the HNN's grows only $2.5\\times$, so the prior matters most where the vector field is hardest to fit locally.","The 3D Kepler results show the prior transfers to six-dimensional phase space: energy drift $8.0\\times$ smaller, angular-momentum drift $4.2\\times$ smaller, and trajectory MSE $5.5\\times$ smaller at $T=10$, with angular momentum not explicitly supervised.","An explicit Störmer–Verlet rollout of a learned non-separable Hamiltonian does not improve over HNN+RK4, implying that structure preservation requires matching the integrator's assumptions (separability or a general symplectic scheme) to the learned model.","On the 1-DOF pendulum, learned surrogates are slower than RK4 with the analytical vector field, so the computational case for HNNs must rest on many-body scaling, which this paper leaves open."],"supporting_citations":[{"why":"Introduces the HNN architecture and symplectic-gradient loss that the paper evaluates under matched conditions.","marker":"[4]"},{"why":"Provides the Adam optimizer used identically to train both models in the protocol.","marker":"[8]"},{"why":"Supplies the symplectic-integration theory behind the Störmer–Verlet diagnostic and bounded-drift expectations.","marker":"[5]"},{"why":"Offers the structure-preserving SympNets baseline relevant to the integrator-compatibility discussion.","marker":"[6]"},{"why":"Defines the Neural ODE vector-field parameterization that Model A resembles.","marker":"[2]"},{"why":"Frames physics-informed machine learning and motivates the architectural-prior comparison.","marker":"[7]"}],"fun_headline_variants":["Hamiltonian prior yields 42x lower drift than matched baseline","Matched integrator isolates Hamiltonian advantage: 42x energy drift cut","Same data, same optimizer: Hamiltonian nets beat feedforward 42x","Energy drift saturates with Hamiltonian net, grows 42x with baseline","Hamiltonian structure, not extra data, drives 42x drift reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the matched protocol isolates the Hamiltonian prior, because the two models do not start from identical initial weights and the Hamiltonian network has a different output head (a single scalar output, 65 fewer parameters), so part of the gap could in principle come from those remaining differences rather than from Hamilton's equations.","fun_headline_variants_meta":{"raw":{"variants":["Hamiltonian prior yields 42x lower drift than matched baseline","Matched integrator isolates Hamiltonian advantage: 42x energy drift cut","Same data, same optimizer: Hamiltonian nets beat feedforward 42x","Energy drift saturates with Hamiltonian net, grows 42x with baseline","Hamiltonian structure, not extra data, drives 42x drift reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00086,"raw_usage":{"total_tokens":3830,"prompt_tokens":1143,"completion_tokens":2687,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":2592}},"tokens_in":759,"tokens_out":2687,"duration_ms":15445,"temperature":1.0,"reasoning_tokens":2592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:10:57.083892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a third architecture under the identical protocol—a scalar-output network whose two-dimensional vector field is produced by multiplying the gradient by a fixed, non-Hamiltonian matrix—and compare long-horizon energy drift; if its drift also saturates near $2\\times10^{-4}$, the reported $42\\times$/$15.8\\times$ ratios would be an effect of scalar parameterization rather than of Hamilton's equations.","supporting_citations":[{"cited_title":"Hairer, C","cited_arxiv_id":null,"evidence_quote":"Supplies the symplectic-integration theory behind the Störmer–Verlet diagnostic and bounded-drift expectations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frames physics-informed machine learning and motivates the architectural-prior comparison."}],"review_version":1}