△ Mechanism confirmed, baseline not beaten 2026

Transport-PDE Predictor for Delayed Neural State Updates

Implementation & benchmark of arXiv:2609.03345 — Traffic Congestion Control for ARZ Model with an Arbitrarily Large Input Delay

Usefulness7/10

Difficulty6/10

Novelty6/10

Dynamics Stability Accuracy Rnn Ssm Optimizer Training dynamics Control theory Dynamical systems Pde Linear algebra Optimization

Source paper: Traffic Congestion Control for ARZ Model with an Arbitrarily Large Input Delay arXiv:2609.03345 ⓘ · analyzed Sep 4, 2026

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 Not peer-reviewed.

Idea description

Add an explicit transport-delay state to a recurrent neural network, state-space model, or learned optimizer whenever actions, gradients, or control inputs arrive after a fixed delay. Use the queued inputs to construct a finite-horizon predictor state and apply the neural transition or controller to that predicted state rather than to the stale state. The design transfers the paper's delay-as-transport-PDE and backstepping-to-stable-target strategy into a differentiable predictor with an observable contraction margin.

Formulas

$$q_t(s,t)+q_s(s,t)=0,\qquad q(0,t)=u(t),\qquad q(\tau,t)=u(t-\tau).$$

$$\hat{x}_{k+d}=A^d x_k+\sum_{j=0}^{d-1}A^{d-1-j}B u_{k+j-d},\qquad u_k=K\hat{x}_{k+d}.$$

$$\rho(A+BK)<1\quad\text{or}\quad (A+BK)^\top P(A+BK)-P\preceq-\alpha P.$$

$$\|x_k\|\leq C\lambda^k\|x_0\|+\gamma\sup_{j<k}\|e_j\|,\qquad \gamma L_e<1.$$

Mathematical statement

Represent a delay of length $\tau$ by the transport equation $q_t(s,t)+q_s(s,t)=0$, for $s\in[0,\tau]$, with boundary condition $q(0,t)=u(t)$; therefore $q(\tau,t)=u(t-\tau)$. For a locally linearized neural state $x_{k+1}=Ax_k+Bu_{k-d}+w_k$, the delay-compensated predictor is $\hat{x}_{k+d}=A^d x_k+\sum_{j=0}^{d-1}A^{d-1-j}B u_{k+j-d}$, where $d$ is the integer delay, $A=\partial f/\partial x$, $B=\partial f/\partial u$, and $w_k$ is model error. Apply $u_k=K\hat{x}_{k+d}$. The target closed-loop matrix is $A+BK$, and exponential stability requires $\rho(A+BK)<1$, or, more robustly, $(A+BK)^\top P(A+BK)-P\preceq-\alpha P$ for some $P\succ0$ and $\alpha>0$. With prediction error $e_k$, an ISS bound has the form $\|x_k\|\leq C\lambda^k\|x_0\|+\gamma\sup_{j<k}\|e_j\|$; the small-gain requirement is $\gamma L_e<1$, where $L_e$ bounds the error feedback gain.

Implementation notes

Integrate this at the hidden-state update of a GRU-free RNN, a diagonal-plus-low-rank SSM, or a learned optimizer whose parameter update is delayed by asynchronous workers. Maintain a FIFO queue containing the last d inputs or gradients. At every step, estimate the local Jacobians A and B of the neural transition f(x,u) using automatic-differentiation Jacobian-vector products; for the first MVP, use a frozen linearization estimated once per minibatch. Compute the predictor with repeated matrix-vector products: set p=x, then apply p <- A@p+B@queued_input[j] for each future slot in chronological order. Feed p, rather than the stale x, into the controller u=pi_theta(p), or use a linear K during the stability experiment. Pseudocode is: enqueue the newest input; p=x; for j=0,...,d-1, p=A@p+B@queued_input[j]; u=controller(p); execute u after the prescribed delay; update the network normally. The transport-PDE part supplies the queue interpretation; A, B, and the residual prediction error are estimated empirically. First test on a 2-layer tanh RNN learning the delayed-copy task and on an SSM trained on sequential MNIST, comparing ordinary delayed feedback with the predictor for d in {1,4,16,64}. Estimate rho(A+BK) by power iteration along trajectories and sweep the controller gain until it crosses one. The predicted signature is a sharp divergence boundary near rho(A+BK)=1, largely independent of d for the compensated model, while the uncompensated model's stable delay range shrinks as d increases. Also inject bounded Gaussian transition noise: state norms should remain bounded below the measured small-gain threshold and grow rapidly when the estimated product gamma times L_e exceeds one.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 9/10

Built a readable delayed linear-system MVP with a FIFO transport queue, finite-horizon predictor, spectral-radius/Lyapunov verification, and fixed-seed comparisons. The transport identity and predictor formula matched numerically to 0 and 1.1e-16 maximum error; the target closed loop had rho=0.825 and a positive Lyapunov matrix with residual maximum eigenvalue -1.0. Across delays 2-32, the stale-state baseline degraded and diverged dramatically for d>=4, while the predictor remained bounded and converged to approximately zero without noise; under transition noise it stayed bounded with RMS 0.112 and 0.449 for noise scales 0.005 and 0.02. This is a clear win in the intended linear, known-model setting, but not yet evidence for nonlinear learned RNNs.

Agent confidence: 9/10

Baseline: RMS final-50 state norm: d=1 3.39e-23, d=2 0.883, d=4 1.49e12, d=8 1.56e14, d=16 1.18e12, d=32 1.95e9; at d=16 with noise sigma=0.005/0.02: 1.17e12/1.14e12
Idea: RMS final-50 state norm: d=1 7.68e-18, d=2 9.38e-18, d=4 1.73e-17, d=8 4.91e-17, d=16 6.69e-17, d=32 4.41e-14; at d=16 with noise sigma=0.005/0.02: 0.112/0.449; target rho(A+BK)=0.8249

Limitations: Only a small known linear plant was tested; no nonlinear RNN, learned Jacobian estimation, training/accuracy task, gain sweep around the stability boundary, GPU execution, or empirical small-gain threshold was evaluated.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2609.03345) Sep 4, 2026

Ahead of the baseline (−8.6%) but not statistically significant (p=0.7344, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): -0 (−8.6%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.7344 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track and fixed delay d=4 were tested. The predictor uses a frozen local linearization rather than learned automatic-differentiation Jacobians; the benchmark showed a small nonsignificant improvement, not a significant win.

How to run: python3 official_stage2.py

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

Stage 2 — Benchmark Sep 4, 2026 latest Failed ✗

Agent confidence: 9/10

Ran the official registered dynamics track with canonical rnn_small, train_model, equal three-point learning-rate sweeps, and 8 paired seeds. The transport predictor slightly improved mean test MSE from 5.152216147052968e-06 to 4.707826093408585e-06, but the paired permutation test was not significant (delta_mean=-4.4439005364438344e-07, p=0.7344), so the canonical verdict is no significant win and system_worked=false. The trained-model mechanism signature was confirmed with mean predicted terminal-state shift 0.21677356958389282.

Baseline: best lr=0.009; mean test MSE=5.152216147052968e-06
Comparison: delta_mean=-4.4439005364438344e-07; p_value=0.7344; verdict=no significant win
Idea: best lr=0.009; mean test MSE=4.707826093408585e-06

Benchmark result