✗ Failed on benchmark 2026

ISS-Gated Positive Neural State Module

Implementation & benchmark of arXiv:2608.13302 — Input-to-state stability of chemical reaction networks with application to molecular computation

Usefulness7/10

Difficulty5/10

Novelty7/10

Dynamics Stability Accuracy Rnn Ssm Training dynamics Optimizer Dynamical systems Control theory Optimization Geometry

Source paper: Input-to-state stability of chemical reaction networks with application to molecular computation arXiv:2608.13302 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained recurrent or neural-ODE hidden state with a positive state driven by reaction-like polynomial flows whose rate vector is modulated by inputs or context. Train the module together with an ISS penalty so bounded gate perturbations produce a bounded hidden-state deviation, preventing long-horizon amplification while retaining nonlinear computation.

Formulas

$$\dot{x}=f(x,u)=\sum_{j=1}^{r}u_j(t)x^{v_{\cdot j}}(v'_{\cdot j}-v_{\cdot j}),\qquad x^{v_{\cdot j}}=\prod_{i=1}^{n}x_i^{v_{ij}}.$$

$$V(x;x^*)=\sum_{i=1}^{n}\left[x_i\log\frac{x_i}{x_i^*}-x_i+x_i^*\right],\qquad \dot V=\sum_i\log\frac{x_i}{x_i^*}\,f_i(x,u).$$

$$\dot V\le -cV+k\|u-u^*\|_2^2\quad\Longrightarrow\quad V(t)\le e^{-ct}V(0)+\frac{k}{c}(1-e^{-ct})\|u-u^*\|_{\infty,[0,t]}^2.$$

$$x=Dz,\qquad \dot z=D^{-1}f(Dz,u),\qquad D=\operatorname{diag}(d_1,\ldots,d_n),\ d_i>0.$$

Mathematical statement

A mass-action network has state x in R^n_{>0}, stoichiometric vectors s_j=v'_{.j}-v_{.j}, rate input u_j(t)>0, and dynamics f_i(x,u)=sum_j s_{ij}u_j x^{v_{.j}}, where x^{v_{.j}}=prod_i x_i^{v_{ij}}. The paper's stability mechanism certifies semiglobal ISS around an equilibrium x* corresponding to nominal rates u*: there exist class-KL beta and class-K functions gamma such that ||x(t)-x*|| <= beta(||x(0)-x*||,t)+gamma(||u-u*||_infty). For a complex-balanced or transformed positive module, use the pseudo-Helmholtz free energy V(x)=sum_i[x_i log(x_i/x_i*)-x_i+x_i*]. Its derivative is dV/dt=grad V(x)^T f(x,u), and the nominal dynamics should satisfy dV/dt <= -alpha(||x-x*||) while rate disturbances obey |dV/dt| <= -c V(x)+k||u-u*||^2. Here c>0 is the nominal decay rate and k is the disturbance sensitivity; the resulting steady-state energy bound is limsup V <= (k/c)||u-u*||_infty^2. Linear-conjugacy-inspired rescaling uses x=Dz with diagonal D>0 and transforms the vector field as zdot=D^{-1}f(Dz,u), allowing the same certificate to be searched in better-conditioned coordinates.

Implementation notes

1. Integration point: use this as a recurrent or state-space block for sequence prediction, world models, or an optimizer state. Let h in R^n_{>0} be the hidden state, let q_t be the feature input, and produce positive rate gates u_t=softplus(Aq_t+b)+epsilon. Implement h_{t+1}=softplus(h_t+Delta*f_theta(h_t,u_t)) or, preferably, integrate the positive flow with two or four RK4 substeps and parameterize h=softplus(z). Include a learned diagonal rescaling D if state coordinates have very different magnitudes. 2. Pseudocode: compute u=softplus(Aq+b)+eps; evaluate f=sum_j u[j]*prod(h**reactant[:,j])*stoich[:,j]; take an RK4 step; compute V=sum(h*log(h/h_star)-h+h_star); estimate Vdot=(V_next-V)/Delta; add loss_iss=lambda_iss*relu(Vdot+c*V-k*||u-u_star||^2)^2. Clip or project rates to a compact interval [u_min,u_max], matching the paper's semiglobal compact-input assumption. 3. Computed from the mechanism: the positive mass-action vector field, free-energy form, and ISS inequality. Estimated empirically: equilibrium h_star, decay c, gain k, and the best diagonal D. Estimate c and k by constrained regression over replayed states: fit -Vdot approximately cV-kd^2, where d=||u-u_star||^2, then validate on held-out disturbances. 4. First cheap experiment: compare vanilla GRU, unconstrained neural ODE, and this module on permuted-MNIST or a noisy Lorenz forecasting task, training with identical parameter counts and testing horizons 1x, 5x, and 20x the training horizon. Inject bounded rate or context perturbations with amplitude a. The predicted signature is an ISS plateau: after transients, V should scale approximately as (k/c)a^2 rather than grow with horizon, and nominal perturbations should decay at slope at least -c in a semilog V plot. Sweep the integration step Delta; instability should occur near the measured condition Delta*c_eff approximately 2 for an explicit Euler variant, while the ISS-trained module should show a substantially larger bounded-stability region than the unconstrained baseline.

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 9/10

Built a readable positive mass-action module with generic reaction stoichiometry, RK4 integration, positivity flooring, free-energy computation, and an ISS residual. The three required mechanism predictions manifested: Euler stability boundary 1.951 versus predicted 2.0, disturbance-energy log-log slope 1.973 versus 2.0, and nominal free-energy decay rate -1.964 versus -2.0; the direct mass-action formula check had absolute error 0.0. The sequence comparison was mixed: the positive module had worse training MSE but better long noisy-horizon MSE than the tanh baseline, so the stability mechanism is supported but a general accuracy win is not established.

Agent confidence: 9/10

Baseline: tanh RNN: train MSE 0.00701; long noisy-horizon MSE 0.00850
Idea: positive mass-action state: train MSE 0.00941; long noisy-horizon MSE 0.00749; Euler boundary 1.951 vs 2.0 predicted; V amplitude slope 1.973 vs 2.0; nominal log-V rate -1.964 vs -2.0; mass-action field error 0.0

Limitations: Only a scalar birth-death mass-action system and a small random-search sequence demo were tested. No trained GRU, multidimensional complex-balanced network, learned diagonal rescaling, explicit ISS-penalty optimization, GPU implementation, or large real sequence benchmark was evaluated.

How to run: python3 iss_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.13302) Sep 2, 2026

Ahead of the baseline (−17.6%) but not statistically significant (p=0.0757, wins 6 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0029
Idea mean: 0.0024
Effect (Δ): -0.0005 (−17.6%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.0757 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±23.7%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 6 configs

Limitations:

Only the built-in dynamics track was tested, with 400 training samples, 200 test samples, 12 epochs, one hidden width, one positive reaction structure, and three idea configurations. No long-horizon rollout metric, learned diagonal rescaling, RK4 integration, multidimensional reaction networks, or GPU/CPU ablation was evaluated.

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 ISS-gated positive neural state module on the structurally matched dynamics track using an end-to-end positive birth/death recurrent state and differentiable ISS penalty. Across 8 paired seeds, test MSE improved from 0.0029277411 to 0.0024121132, a 17.6% reduction with 6/8 wins, but the paired permutation p-value was 0.0757, so there was no significant win. The trained states stayed positive and finite, but observed mean Vdot was positive (1.7622), so the quantitative ISS decay prediction was not confirmed.

Baseline: rnn_small; best lr=0.003, weight_decay=0.0001; mean test MSE=0.0029277411
Delta mean: -0.0005156279
Idea: Positive ISS recurrent state; lr=0.003, weight_decay=0.0001, lambda_iss=0.03; mean test MSE=0.0024121132
Idea wins: 6
N pairs: 8
Paired permutation p: 0.0757
Relative improvement pct: 17.614

Benchmark result