✗ Failed on benchmark 2026

PIPO-PITO bounded recurrent gain

Implementation & benchmark of arXiv:2608.01699 — On input-output persistency and the interconnection of positive nonlinear systems

Usefulness7/10

Difficulty5/10

Novelty7/10

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

Source paper: On input-output persistency and the interconnection of positive nonlinear systems arXiv:2608.01699 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Build a positive continuous-depth RNN or state-space layer in which a nonnegative recurrent-input gain is generated by a PITO controller. If sustained large gain produces sustained large hidden-state output through a PIPO plant, the controller automatically decreases the gain, preventing runaway recurrent dynamics without requiring a globally tiny fixed gain. The construction predicts a quantitative attenuation threshold and exponential decay rate when the hidden output stays above that threshold.

Formulas

$$\dot{x}=f(x,u),\qquad y=h(x),\qquad u(t)=w(t),\qquad v(t)=y(t).$$

$$u(t)\geq U(\kappa)\ \forall t\geq 0\quad\Longrightarrow\quad y(t)\geq \kappa\ \forall t\geq T(\kappa).$$

$$\dot{z}=b-a\,y(t)z,\qquad w=z,\qquad y(t)\geq V\quad\Longrightarrow\quad w(t)\leq w(0)e^{-aVt}+\frac{b}{aV}.$$

$$V(\varepsilon)=\frac{2b}{a\varepsilon},\qquad T(\varepsilon,w_0)=\max\left\{0,\frac{\log(2w_0/\varepsilon)}{aV(\varepsilon)}\right\}.$$

Mathematical statement

The plant has nonnegative state x in R^n, nonnegative input u, and output y=h(x)>=0, with dynamics dx/dt=f(x,u). It satisfies PIPO if functions U(kappa) and T(kappa) exist such that u(t)>=U(kappa) for all t>=0 implies y(t)>=kappa for all t>=T(kappa), for every initial state. The controller has dz/dt=g(z,y), output w=k(z)>=0, and satisfies PITO if y(t)>=V(epsilon) for all t>=0 implies w(t)<=epsilon after a time depending on epsilon and the initial output. Interconnect with u=w and v=y. The paper's theorem states that forward completeness, PIPO, PITO, and a one-sided affine controller-output growth condition imply bounded u=w. A concrete PITO controller is w=z and dz/dt=b-a y z, with a,b>0. When y>=V, z(t)<=z(0) exp(-a V t)+b/(a V). Choosing V(epsilon)=2b/(a epsilon) yields w<=epsilon after a computable transient time.

Implementation notes

1. Integration point: replace the fixed recurrent input gain in a continuous-depth RNN or positive state-space model by a scalar or channelwise state w. Use a positive plant such as dx/dt=(-D+A)x+B w, y=q^T x, where A,B,q are nonnegative, D-A is an M-matrix or otherwise makes the plant forward complete, and x is initialized nonnegative. Inject w into the recurrent state, and use the controller dz/dt=b-a y z with w=z. The measured output y can be the mean, sum, or a selected nonnegative projection of hidden activations. 2. Pseudocode: initialize x>=0 and w>=0; at each substep compute x=relu(x+dt*((-D+A)@x+B*w)); compute y=q@x; then update w=relu(w+dt*(b-a*y*w)); use x as the recurrent representation. Run this update once per token or continuous-depth substep. 3. Computed versus estimated: a,b,D,A,B,q are selected hyperparameters. Estimate the PIPO curve by holding w(t)=u0 constant across several initial states and recording the smallest time T at which y>=kappa. The global theorem assumptions must be stress-tested numerically. Fit the observed controller response to w0 exp(-a V t)+b/(a V). 4. First cheap experiment: use a 16-dimensional positive linear SSM on a synthetic delayed-copy and adding task, comparing fixed-gain, GRU, and PIPO-PITO models at equal parameter count. Sweep sequence length and initial recurrent gain. The quantitative prediction is that whenever y remains above V(epsilon)=2b/(a epsilon), w falls below epsilon at the predicted time within approximately 20 percent. A plot of log(w-b/(aV)) versus time should have slope -aV. With feedback disabled, increasing the fixed gain beyond the discrete plant stability boundary rho(I+dt(-D+A))>=1 should produce activation growth, while the regulated model should keep w bounded.

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 and numerically verified a scalar positive PIPO plant with the PITO controller w'=b-a y w. Measured exponential decay rates matched the predicted aV within 0.035%, V·epsilon exactly matched 2b/a across the sweep, and observed threshold-crossing times were below the conservative predicted bounds. Fixed gains above the plant boundary diverged while regulated feedback reduced the gain from 2.0 to 0.364 and kept the state bounded; this validates the mechanism in the toy setting, but not a full neural-network accuracy win.

Agent confidence: 8/10

Baseline: Fixed-gain plant: gamma=0.2 and 0.39 remained bounded (final x=0.100 and 1.149); gamma=0.6 and 0.8 diverged to 1.83e6 and 1.22e13 by t=80. The marginal gamma=0.4 case grows linearly but did not exceed the finite-horizon divergence criterion.
Idea: PITO constant-output sweep: maximum relative error in predicted decay rate was 0.000350 (0.035%); V*epsilon=1.076923 for every epsilon, exactly equal to 2b/a; maximum observed-minus-predicted crossing time was -0.00499. Closed-loop at nominal gamma=0.8 had max x=1.444, final x=0.550, and final w=0.364.

Limitations: Only a scalar continuous-time toy plant was tested; no 16-dimensional delayed-copy/adding task, GRU comparison, parameter-matched neural model, FLOP/loss benchmark, channelwise controller, or GPU experiment was run. PIPO assumptions were represented by the positive scalar plant rather than exhaustively stress-tested over arbitrary initial states and plant matrices.

How to run: python3 pipo_pito_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.01699) Aug 31, 2026

No significant difference vs. the tuned baseline (+10.2%, p=0.4142, wins 1 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: positive_rnn_fixed_gain_vs_pito
Paired seeds: 8
Baseline mean: 0.0003
Idea mean: 0.0004
Effect (Δ): 0 (+10.2%; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.4142 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±28.0%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 9 configs

Limitations:

Only the registered built-in dynamics track was tested. No other tracks were structurally matched to the stability/control idea. The selected trained-model run did not yield enough valid high-output decay traces for quantitative signature confirmation; the lower-learning-rate exploratory setting produced a relative signature error of approximately 99.86%.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

The PITO recurrent-gain system was evaluated against the same positive recurrent architecture with a fixed gain on the registered dynamics/control track. The tuned baseline reached mean MSE 0.0003413614 versus 0.0003760458 for the idea; paired delta was +0.0000346845 with permutation p=0.41415, so there was no significant win. The trained-model mechanism signature was not confirmed.

Baseline: Best config lr=0.01, gain=1.5; full 8-seed mean MSE 0.00034136135218432173.
Comparison: delta_mean=0.0000346844954037806, idea_wins=1/8, p_value=0.41415, verdict=no significant win.
Idea: Best config lr=0.01, a=1.0, b=0.2, initial gain=1.0; full 8-seed mean MSE 0.00037604584758810233.

Benchmark result