✗ Failed on benchmark 2026

Input-Aware Contracting Neural ODE

Implementation & benchmark of arXiv:2608.09404 — Adaptive Stability-Constrained Neural Differential Equations for Controlled Dynamical Systems with Unknown Inputs

Usefulness8/10

Difficulty6/10

Novelty6/10

Dynamics Stability Accuracy Generalization Optimizer Training dynamics Ssm Rnn Control theory Dynamical systems Differential geometry Optimization Linear algebra

Source paper: Adaptive Stability-Constrained Neural Differential Equations for Controlled Dynamical Systems with Unknown Inputs arXiv:2608.09404 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Train a neural vector field together with a positive-definite metric $M_\phi(x,u)$ that certifies local contraction at a prescribed rate. The contraction penalty must include the total derivative of the input-dependent metric, so rapidly changing controls are treated as a source of geometry variation rather than incorrectly claiming stability from a frozen metric.

Formulas

$$M_\phi(x,u)=L_\phi(x,u)L_\phi(x,u)^\top+\epsilon I,\qquad V(\delta x,x,u)=\delta x^\top M_\phi(x,u)\delta x.$$

$$\dot M_\phi=\frac{\partial M_\phi}{\partial x}f_\theta(x,u)+\frac{\partial M_\phi}{\partial u}\dot u,$$

$$\dot M_\phi+A^\top M_\phi+M_\phi A+2\lambda M_\phi\preceq0,\qquad A=\frac{\partial f_\theta}{\partial x}.$$

$$\mathcal L_{\mathrm{contr}}=\mathbb E\left[\operatorname{softplus}\left(\lambda_{\max}\left(M^{-1/2}SM^{-1/2}\right)\right)\right],\quad S=\dot M+A^\top M+MA+2\lambda M+\alpha M+\beta MBB^\top M+\gamma MEE^\top M.$$

Mathematical statement

The learned nominal dynamics are $\dot x=f_\theta(x,u)$, where $x\in\mathbb R^n$ is the state, $u\in\mathbb R^m$ is the measured input, and $\theta$ are vector-field parameters. Define a positive-definite metric $M_\phi(x,u)=L_\phi(x,u)L_\phi(x,u)^\top+\epsilon I$, where $L_\phi$ is a lower-triangular neural network, $\epsilon>0$, and $\phi$ are metric parameters. For an infinitesimal displacement $\delta x$, use $V=\delta x^\top M_\phi\delta x$. Let $A=\partial f_\theta/\partial x$ and $B=\partial f_\theta/\partial u$. Along an input trajectory, the total metric derivative is $\dot M=(\partial M/\partial x)f_\theta+(\partial M/\partial u)\dot u$. The local contraction certificate at rate $\lambda>0$ is $\dot M+A^\top M+MA+2\lambda M\preceq0$. For input and disturbance mismatch, with disturbance injection matrix $E$, a sufficient robust condition adds $\alpha M+\beta MBB^\top M+\gamma MEE^\top M$, where $\alpha,\beta,\gamma>0$ are design weights. A negative largest generalized eigenvalue predicts exponential decay for matched trajectories and a bounded input-to-state separation for mismatched trajectories.

Implementation notes

1. Integration point: replace an unconstrained NODE or continuous-time state-space model with two networks, $f_\theta(x,u)$ and a metric-factor network $L_\phi(x,u)$. Add the contraction loss to the trajectory prediction loss, evaluating it on states from training trajectories and on randomly perturbed rollout states. Estimate $\dot u$ from adjacent control samples using centered finite differences; for piecewise-constant controls, use zero inside intervals and evaluate separately at switching points. 2. Pseudocode: for each minibatch, compute $f=f_\theta(x,u)$, $A=\partial f/\partial x$, $B=\partial f/\partial u$, and $M=LL^\top+\epsilon I$. Use automatic differentiation to obtain $\partial_xM$ and $\partial_uM$, then form $\dot M=(\partial_xM)f+(\partial_uM)\dot u$. Construct $S$ from the displayed robust inequality, estimate its largest generalized eigenvalue through $M^{-1/2}SM^{-1/2}$, and optimize $\mathcal L_{\mathrm{prediction}}+\rho\mathcal L_{\mathrm{contr}}$. Increase $\rho$ when the violation fraction exceeds a target and decrease it when violations remain rare. 3. Computed from the mechanism are the vector-field Jacobians, metric derivative, and generalized eigenvalue. Estimated empirically are the control derivative, state sampling distribution, and disturbance magnitude; initially set $E=0$, then test known or conservative disturbance channels. 4. First experiment: train a two-state forced Duffing NODE on short trajectories, comparing an unconstrained NODE, a fixed Euclidean metric $M=I$, and the adaptive metric. Sweep the learning rate and evaluate long rollouts under held-out forcing. The predicted signature is that the constrained model's sampled maximum generalized eigenvalue $\mu_{\max}=\lambda_{\max}(M^{-1/2}SM^{-1/2})$ becomes negative, while the unconstrained model remains positive in some regions. For identical inputs, the logarithm of the metric distance between nearby rollouts should have slope no greater than $-\lambda$ until approximation error dominates. Under input mismatch, separation should approach a plateau proportional to the mismatch norm rather than grow exponentially. Removing the $(\partial M/\partial u)\dot u$ term should selectively produce certificate violations and worse rollouts for rapidly varying controls.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Inconclusive

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built an input-aware contracting Neural ODE prototype with a positive-definite learned metric and total derivative dM/dt=(∂M/∂x)f+(∂M/∂u)u_dot. The scalar checks confirmed three predictions essentially exactly: boundary q*=0.8, sensitivity dμ/dq=1.5, and zero input-rate effect when the metric is input-independent. The adaptive model reduced mean certificate eigenvalue and violation fraction, but its worst-case violation remained large, so this is a promising mechanism signal rather than a complete stability certificate.

Agent confidence: 8/10

Baseline: Unconstrained: mean_mu=1.675, max_mu=2.295, violation_fraction=1.000; fixed M=I: mean_mu=0.748, max_mu=1.589, violation_fraction=1.000.
Idea: Adaptive input-aware metric with total derivative: mean_mu=-0.402, max_mu=22.465, violation_fraction=0.438. Analytic predictions: q boundary 0.800 predicted/observed, dmu/dq 1.500 predicted/observed, k=0 variation 0.

Limitations: The mini-experiment uses a small short-horizon Duffing-inspired vector-field fitting task rather than full NODE trajectory optimization, omits robust B/E disturbance terms, does not separately train a frozen-metric neural model, and does not test long rollouts or the input-mismatch plateau. The learned metric still has substantial worst-case certificate violations.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.09404) Sep 2, 2026

Ahead of the baseline (−4.5%) but not statistically significant (p=0.3924, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.085
Idea mean: 0.0812
Effect (Δ): -0.0038 (−4.5%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.3924 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±10.9%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

The benchmark used a reduced but equal-budget dynamics dataset of 240 training and 120 test samples with 4 epochs because higher-order autograd through the GRU was expensive. The local certificate is a scalar approximation applied to the final theta/control slice of rnn_small rather than a full two-state NODE vector-field certificate; robust disturbance terms were omitted. The analytical scalar derivative check passed, but it is not counted as NN-scale mechanism confirmation; a baseline trained-model certificate aggregate was not retained in the final report.

How to run: python3 bench_experiment.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the input-aware contracting metric as a training-loss intervention on the structurally matched dynamics/rnn_small benchmark, including a positive-definite learned metric and total derivative contribution from state and control rate. Across 8 paired seeds, test MSE improved slightly from 0.0850381 to 0.0812273, but the permutation p-value was 0.3924, so there was no significant win. The trained certificate remained violated on nearly all probe points, and the NN-scale quantitative mechanism prediction was not confirmed.

Baseline: dynamics/rnn_small, tuned lr=0.006, test MSE mean=0.0850381, std=0.0185276
Idea: input-aware metric penalty, lr=0.006, rho=0.05, test MSE mean=0.0812273, std=0.0218786; paired delta=-0.00381088, p=0.3924; idea wins=4/8

Benchmark result