{"id":"4479f136-a97a-4d5d-8550-8e5e8e4d6c3c","arxiv_id":"2411.13914","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ICODE, a control-affine neural ODE that feeds external inputs directly into the dynamics, predicts trajectories better than NODE, ANODE, and CDE on simulated physical benchmarks, but its contraction guarantee is not enforced or verified.","lead":"This paper introduces ICODE, a neural network model that learns physical systems by feeding measured external inputs directly into the dynamics. It is tested on several simulated systems and beats three existing neural ODE methods, especially when inputs switch abruptly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's contraction condition does not imply fixed-point convergence for time-varying inputs; since the experiments use time-varying u, the central guarantee is unsupported and false in a simple scalar counterexample.","rationale":"The reader's weakest_assumption focused on the benchmark systems being affine in u and on input availability. That is an applicability concern. However, the reader's rationale also flagged that the contraction guarantee is not actually delivered and that the paper conflates contraction of state trajectories with convergence of training. My concern is sharper: even when every applicability assumption is satisfied (system affine in u, u measured at all times), Theorem 1 as stated does not yield the advertised fixed-point convergence. Contraction of a non-autonomous system only ensures convergence of trajectories to each other, not to an equilibrium. The counterexample ẋ = -x + sin(t) satisfies condition (5) but has no fixed-point limit. This directly undermines the abstract's central theoretical claim and one of the paper's four listed contributions. The empirical results may still be plausible and useful, and the theoretical claim could be repaired by replacing 'converge to a fixed point' with 'converge to a common trajectory' (plus a verification condition), which is why I keep the reader's CONDITIONAL verdict rather than moving to REJECT. The concrete check above is cheap and definitive: it isolates whether the fixed-point conclusion follows from condition (5), independent of training details and benchmarks. If the authors amend the theorem statement accordingly, the paper's core method remains valuable; if they insist on the fixed-point formulation, the central claim is false.","tokens_in":36411,"tokens_out":7266,"duration_ms":82388,"concrete_test":"Settle the concern analytically with a scalar instance of Eq. (4): set f(x) = -x, k(x) = 1, u(t) = sin(t). Verify condition (5) with L(x) = 1 and c = 1, since R = -1. Solve exactly: x(t) = e^{-t} x(0) + ∫_0^t e^{-(t-s)} sin(s) ds, which tends to (sin t - cos t)/2 as t → ∞, not a fixed point. This refutes the claimed implication 'contraction ⇒ fixed point' for time-varying inputs. As a numerical check, construct or train an ICODE on ẋ = -x + u(t), verify R + R^T < 0 during the trajectory, integrate over a long horizon, and observe that the state does not converge to a constant when u is sinusoidal, whereas it does when u is constant.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing issue is the inference from Theorem 1 to the abstract's guarantee that 'system trajectories of the trained model converge to a fixed point.' In Eq. (4), u(t) is a time-varying external signal, so the vector field is non-autonomous. Contraction under condition (5) gives incremental stability: any two solutions satisfy |δx(t)| → 0, meaning all trajectories converge to a common trajectory x*(t). It does not imply x*(t) is constant. The proof itself stops at 'δy exponentially converges to 0', which is exactly the incremental statement. A fixed-point conclusion requires u(t) → constant or an additional argument that the limit trajectory is an equilibrium. Since the experiments use switching, sinusoidal, and piecewise inputs, the advertised guarantee does not follow even if (5) were enforced. A concrete counterexample is ẋ = -x + sin(t), which has the ICODE form, satisfies (5) with L=1 (R=-1), yet all solutions tend to a sinusoidal steady state, not a fixed point. Thus the central theoretical claim is false as stated; only a weaker 'all solutions converge to a common trajectory' claim is supported. The paper also never verifies (5) on trained networks, so even this weaker guarantee is not checked in practice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Input Concomitant Neural ODEs (ICODEs), a neural-ODE architecture that is affine in the external input u, with separate subnetworks for the drift and input-coupling vector fields. The authors claim that a contraction condition, given in Theorem 1, guarantees that trajectories of the trained model converge to a fixed point regardless of initial conditions, and they support the method with prediction experiments on seven physical systems (single-link robot, DC-DC converter, rigid body, Rabinovich-Fabrikant equation, glycolytic-glycogenolytic pathway, swing equation, heat conduction), comparing against NODE, ANODE, and CDE variants, including noisy-input and scaling studies.","tokens_in":36637,"tokens_out":5298,"duration_ms":56806,"significance":"If the empirical claims are correct, ICODE is a useful and simple input-aware neural-ODE class, and the breadth of benchmarks plus the public code are assets. The core advertised theoretical contribution, however, is not supported: contraction of trajectories does not imply convergence to a fixed point when the input is time-varying, and condition (5) is never verified or enforced on trained models. The experimental comparison is also marred by internally inconsistent metric values (RMSE smaller than MAE in Tables I and VII), which casts doubt on the quantitative superiority claims as currently reported.","major_comments":[{"comment":"The central guarantee is stated as convergence to a fixed point, but Theorem 1 only establishes incremental stability. In Eq. (4), u(t) is a time-varying external signal, so the vector field is non-autonomous. Condition (5) gives exponential convergence of δy to 0, i.e., any two solutions converge to a common trajectory x*(t); it does not imply that x*(t) is constant. A concrete counterexample is ẋ = -x + sin(t), which has the ICODE form with f0(x) = -x, g(x) = 1, satisfies (5) with L = 1, yet every solution converges to a sinusoidal steady state rather than a fixed point. Thus the abstract's claim that 'system trajectories of the trained model converge to a fixed point, regardless of initial conditions' is false for the time-varying inputs used throughout the experiments. I ask the authors to either prove the fixed-point claim under an explicit extra assumption (e.g., u(t) → constant) or replace 'fixed point' with 'common trajectory' throughout the abstract, introduction, and conclusion.","section":"Abstract and Section II (Theorem 1)"},{"comment":"Theorem 1 is not connected to the trained models: the paper never constructs the metric M(x), never checks whether condition (5) holds for any trained ICODE, and imposes no regularization or parameterization that would enforce contraction. The theorem is a restatement of the standard Lohmiller-Slotine criterion applied to the Jacobian of (4), and the proof explicitly defers to [27]. As a result, the statement in the introduction that 'theoretical guarantees are provided for the contraction property of ICODEs' overstates what is demonstrated. I request either (a) an algorithmic way to enforce or certify (5) on trained networks, with verification reported in the experiments, or (b) a clearly stated weakening of the contribution to 'a sufficient condition that holds for some ICODE instances, not verified here.'","section":"Section II, condition (5) and Section III"},{"comment":"The reported RMSE values are smaller than the corresponding MAE values in several rows, which is impossible for any error sequence: RMSE ≥ MAE by Jensen's inequality. For example, Table I (single-link robot, Δu=0) reports RMSE = 0.027 and MAE = 0.065 for ICODE, and Table VII (single-link robot, ku=0) reports RMSE = 0.048 and MAE = 0.18. These entries strongly suggest that what is labeled 'RMSE' is actually mean squared error. If so, all quantitative comparisons in these tables need to be recomputed, re-labeled, or have their units clarified; the current reporting prevents the reader from assessing the claimed superior performance.","section":"Tables I and VII"},{"comment":"The text says contraction 'refers to the convergence of the models to the theoretical optimal solution throughout the learning process, despite variations in the initial conditions of the training processes.' This conflates two distinct notions: (i) trajectory convergence of a given dynamical system and (ii) convergence of the training algorithm to an optimal parameter set. Theorem 1 concerns only (i). This conflation is repeated in the 'Note to Practitioners' and in the contribution list, and it should be corrected to avoid implying that contraction of the learned vector field guarantees convergence of the training procedure.","section":"Section II, paragraph before Theorem 1"}],"minor_comments":[{"comment":"The last sentence of Appendix A ends with the fragment 'significantly weaker.' after discussing ANODE's test-set performance; this should be completed or merged with the preceding sentence.","section":"Appendix A"},{"comment":"In Eq. (3), the activation functions are listed as h0, ..., h_{α−1}, s0, s_{β−1}, which omits s1 through s_{β−2}; the notation should be made consistent, such as s0, ..., s_{β−1}.","section":"Section II, Eq. (3)"},{"comment":"The units 'Wb' (weber) are used for control torque, switch signal, and other inputs; these quantities are not magnetic fluxes. The units should be corrected (e.g., N·m for torque in the robot example, dimensionless for the converter switch, etc.) or removed where the input is dimensionless.","section":"Figures 2, 4, 10"},{"comment":"The parameter 'q = 3.5 × 10^{-4}' is introduced without explanation; if this is a damping coefficient or other physical parameter, it should be named and its units given.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theoretical selling point is untenable as stated because contraction under time-varying inputs does not imply fixed-point convergence; the counterexample ẋ = -x + sin(t) is simple and within the paper's own modeling framework. The contraction condition is also never verified on trained networks, so the guarantee is not operational. The empirical study is broad, but the RMSE/MAE inconsistencies in Tables I and VII suggest metric mislabeling and require a full re-evaluation of the quantitative comparisons. I recommend major revision rather than rejection because the empirical contribution could be salvaged if the theoretical claims are corrected and the metrics are fixed; however, if the authors cannot verify condition (5) or provide a valid fixed-point argument, the central claimed contribution would need to be substantially reduced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for its empirical side. ICODE is a clean, simple idea: a neural ODE with an explicit control-affine input term, dx/dt = sum_i f_i(x) + sum_j k_j(x) u_j, and the authors test it on a wide range of simulated physical systems with code and ablations. The experiments are generally careful, with noise robustness, input-span studies, and scaling behavior. If you need a plug-and-play way to include measured inputs in a neural ODE, this is a reasonable recipe.\n\nThe soft spot is the theory, and it is load-bearing. The abstract and introduction claim that the contraction condition guarantees trajectories converge to a fixed point. Theorem 1 and its proof actually establish incremental stability: any two solutions converge to each other, i.e., to a common trajectory. For time-varying u, that common trajectory is generally not constant. The counterexample xdot = -x + sin(t) is in ICODE form, satisfies condition (5) with L = 1, yet no solution tends to a fixed point. Since all experiments use time-varying or switching inputs, the advertised guarantee does not follow even if (5) were enforced, and the paper never verifies (5) on trained networks. The text also conflates trajectory contraction with training convergence in Section II, which compounds the confusion.\n\nThere are two smaller issues. The paper does not cite the standard control-affine system identification and control literature, which uses exactly this form for decades, so the novelty claim is overstated. And the comparisons lack a natural baseline: a control-affine NODE with a single drift network and a single input network. That baseline would isolate the benefit of the multi-subnetwork parameterization, which the present comparisons against CDE, NODE, and ANODE do not do. Some tables (Table I, VII, VIII) also lack error bars, though other tables include them.\n\nOverall, the empirical method is plausible and the limitations section is honest about input-independence degradation. The problem is the mismatch between the theory and the claims. A revision that restates the guarantee as incremental stability, verifies or enforces the contraction condition, and adds the missing baseline would make the paper solid. I would send it to review, but the reviewers should be told to focus on the fixed-point claim.","headline":"The empirical ICODE method is plausible and worth a look, but the advertised contraction-to-fixed-point theorem only proves incremental stability for time-varying inputs, so the central theoretical claim is false as stated.","tokens_in":37201,"tokens_out":1545,"would_cite":false,"duration_ms":18661,"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":"Feeding external inputs directly into neural ODEs improves prediction on physical systems, with a contraction guarantee for fixed-point convergence.","keywords":["ICODE","neural ODE","external input","dynamical systems","contraction analysis","physical dynamics","affine-in-input modeling","input coupling"],"falsifier":"Train ICODE and a standard NODE on a system where the true input enters nonlinearly, such as $\\dot{x} = f(x) + g(x)u^2$ with $u$ switching piecewise, and compare prediction error on held-out input shapes; if ICODE loses its advantage while the paper's mechanism requires affine-in-$u$ structure, the central claim would be falsified. Alternatively, inspect a trained ICODE with trajectories observed not to converge despite the Corollary 1 condition holding, which would break the contraction guarantee.","tokens_in":36185,"feed_emoji":"⚙️","tokens_out":6308,"duration_ms":63867,"temperature":0.7,"pith_summary":"This paper introduces Input Concomitant Neural ODEs (ICODEs), a class of neural ODEs that write the external input u explicitly into the vector field through an affine coupling term, instead of treating u as a hidden parameter to be inferred. The authors claim that this inductive bias matters: on seven physical systems—including a single-link robot, a DC-to-DC converter, rigid-body motion, the Rabinovich-Fabrikant equation, a metabolic pathway, a power-grid swing equation, and heat conduction—ICODE achieves the best prediction accuracy under both smooth and atypical (piecewise or switching) inputs. They also provide sufficient conditions under which the learned model is contracting, meaning trajectories converge to a fixed point regardless of initial conditions. A sympathetic reader would take the central claim to be that explicit, affine input information makes neural ODEs both more accurate and more stable for input-driven physical dynamics.","feed_headline":"Input-aware neural ODEs beat standard models on physical dynamics","feed_subtitle":"A contraction guarantee keeps learned trajectories converging to a fixed point under any initial condition.","key_machinery":"The central object is the affine-in-input ICODE vector field $\\dot{x} = \\sum_i f_i(x) + \\sum_j k_j(x)u_j$, in which the input $u$ multiplies state-dependent neural networks $k_j$. This is the input-concomitant coupling term; its role is to put the measured external input on the same footing as the state, so that the model does not have to infer $u$ from data or differentiate it in time. The contraction guarantee rests on the metric condition of Theorem 1: an invertible $L$ defining $M = L^\\top L$, with the symmetrized pullback of the Jacobian bounded above by $-cI$; Corollary 1 gives the simpler sufficient condition that the largest eigenvalue of the symmetrized Jacobian of the whole vector field stays below $-c$.","core_discovery":"ICODEs are defined by $\\dot{x} = f_0(x) + g(x)u$, where both $f_0$ and the columns of $g$ are neural networks, so the input enters the dynamics in the same multiplicative way that control inputs typically enter physical systems. The paper proves that if there is a uniformly positive definite metric $M = L^\\top L$ such that the symmetrized generalized Jacobian $R$ satisfies $(R + R^\\top)/2 \\le -cI$, then the model is contracting: any two trajectories converge exponentially to each other, and the system converges to a fixed point independent of the initial condition. The experimental section argues that this structure, plus the absence of a derivative on $u$ (unlike neural CDEs), lets ICODE capture the coupling between state and input, giving consistently lower prediction errors than NODE, ANODE, and CDE, particularly when $u$ switches abruptly or is piecewise.","pith_inferences":["The proof of Theorem 1 is not carried out in the paper; it is attributed to the contraction analysis in [27], so the convergence guarantee inherits the assumptions of that external contraction-analysis result.","A natural next step the paper leaves open is to replace the affine-in-$u$ ansatz with a more general $\\dot{x} = f(x,u)$, which the limitations section names as future work; one could test whether learned coupling terms can capture quadratic or logarithmic input dependencies.","The contraction conditions are checkable during training; an implementation could monitor the spectral bound as a regularizer, a use not explored in the paper.","The heat-conduction experiments suggest ICODE also serves for spatiotemporal PDEs when boundary or source terms are treated as inputs; a testable extension would be to compare against physics-informed baselines on the same PDE."],"forward_implications":["If ICODE is right, then for systems that are affine in a measured input, a neural ODE can learn the state-input coupling directly, yielding accurate multi-step prediction even when the input switches at discontinuities within the test horizon.","The contraction conditions give a training-independent stability certificate: under Theorem 1 or Corollary 1, predictions converge to the same fixed point from any initial condition, which is useful in safety-critical settings.","Because the input enters as a multiplier rather than through a derivative, ICODE avoids the numerical amplification that piecewise inputs cause in neural CDEs, so performance holds as the input slope steepens.","Scaling experiments indicate that ICODE benefits from larger width and depth, with training and validation loss both decreasing, suggesting the architecture is useful beyond small benchmark models.","When the input span is zero or the input is independent of the state, ICODE reduces to standard NODE behavior, as the paper explicitly notes."],"supporting_citations":[{"why":"Defines the neural ODE baseline that ICODE extends and that all comparisons are measured against.","marker":"[5]"},{"why":"Baseline that augments the state space; comparison target especially for overfitting behavior.","marker":"[6]"},{"why":"Baseline that incorporates input via path integration, contrasted with ICODE's direct affine coupling.","marker":"[14]"},{"why":"Supplies the contraction-analysis framework whose metric condition Theorem 1 restates and whose proof is used.","marker":"[27]"},{"why":"Earlier contraction guarantee for a NODE variant, framing ICODE's contraction contribution.","marker":"[17]"},{"why":"Contraction-metrics approach to learning stable dynamics, related to the paper's guarantee.","marker":"[18]"}],"fun_headline_variants":["ICODE: neural ODEs that use real-time input, not hidden params","Input-concomitant neural ODEs out-predict on physical dynamics","Neural ODEs with explicit input coupling ensure contraction","ICODE: modeling physical systems with explicit inputs and a contraction guarantee","Input-aware neural ODEs converge to fixed point, beat CDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the physical system is affine in the measured input $u$ and that $u$ is available accurately at every training and prediction time; if $u$ enters nonlinearly, is missing, or is heavily corrupted, the ICODE advantage can disappear.","fun_headline_variants_meta":{"raw":{"variants":["ICODE: neural ODEs that use real-time input, not hidden params","Input-concomitant neural ODEs out-predict on physical dynamics","Neural ODEs with explicit input coupling ensure contraction","ICODE: modeling physical systems with explicit inputs and a contraction guarantee","Input-aware neural ODEs converge to fixed point, beat CDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1308,"prompt_tokens":970,"completion_tokens":338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":586,"tokens_out":338,"duration_ms":3052,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:45:04.508696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ICODE and a standard NODE on a system where the true input enters nonlinearly, such as $\\dot{x} = f(x) + g(x)u^2$ with $u$ switching piecewise, and compare prediction error on held-out input shapes; if ICODE loses its advantage while the paper's mechanism requires affine-in-$u$ structure, the central claim would be falsified. Alternatively, inspect a trained ICODE with trajectories observed not to converge despite the Corollary 1 condition holding, which would break the contraction guarantee.","supporting_citations":[{"cited_title":"Neural ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Defines the neural ODE baseline that ICODE extends and that all comparisons are measured against."},{"cited_title":"Augmented neural ODEs,","cited_arxiv_id":null,"evidence_quote":"Baseline that augments the state space; comparison target especially for overfitting behavior."},{"cited_title":"Neural controlled differential equations for irregular time series,","cited_arxiv_id":null,"evidence_quote":"Baseline that incorporates input via path integration, contrasted with ICODE's direct affine coupling."},{"cited_title":"Robust classification using contractive Hamiltonian neural ODEs,","cited_arxiv_id":null,"evidence_quote":"Earlier contraction guarantee for a NODE variant, framing ICODE's contraction contribution."},{"cited_title":"Learning stabilizable dynamical systems via control contraction metrics,","cited_arxiv_id":null,"evidence_quote":"Contraction-metrics approach to learning stable dynamics, related to the paper's guarantee."}],"review_version":1}