△ Mechanism confirmed, baseline not beaten 2026

Spectral-Edge Criticality Controller

Implementation & benchmark of arXiv:2608.26279 — Critical Properties and Glass Transitions in Randomly Coupled Fields

Usefulness8/10

Difficulty5/10

Novelty6/10

Dynamics Stability Accuracy Generalization Rnn Ssm Training dynamics Initialization Optimizer Statistical mechanics Random matrix Spectral theory Dynamical systems

Source paper: Critical Properties and Glass Transitions in Randomly Coupled Fields arXiv:2608.26279 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Use the spectral edge of the effective recurrent Jacobian as an explicit control variable. Scale the recurrent coupling so that its largest effective eigenvalue remains a chosen distance below marginal stability, preserving long memory without allowing exploding states or gradients. The mechanism predicts a sharp change in correlation time and gradient persistence when the estimated edge crosses the critical value.

Formulas

$$\rho(\lambda)=\frac{\sqrt{4\sigma^2-\lambda^2}}{2\pi\sigma^2},\qquad \lambda_+=2\sigma$$

$$g_c\lambda_+(J)=1,\qquad g_c=\frac{1}{\lambda_+(J)}\approx\frac{1}{2\sigma}$$

$$C(t)\propto(g\lambda_+)^t=\exp\!\left[-\frac{t}{\xi}\right],\qquad \xi=-\frac{1}{\log(g\lambda_+)}\approx\frac{1}{1-g\lambda_+}$$

$$g_{t+1}=\operatorname{clip}\left(g_t\exp\left[\alpha\left((1-\varepsilon)-g_t\widehat{\lambda}_{+,t}\right)\right],g_{\min},g_{\max}\right)$$

Mathematical statement

For a symmetric Wigner coupling matrix W with entry variance $\sigma^2/n$, the limiting eigenvalue density is $\rho(\lambda)=\sqrt{4\sigma^2-\lambda^2}/(2\pi\sigma^2)$ on $[-2\sigma,2\sigma]$, with upper spectral edge $\lambda_+=2\sigma$. Consider linearized recurrent dynamics $\delta h_{t+1}=gJ_t\delta h_t$, where $h_t\in\mathbb{R}^n$ is the hidden state, $g>0$ is a controlled gain, and $J_t$ is the local Jacobian excluding this gain. In the frozen symmetric approximation, marginal stability occurs at $g\lambda_+(J)=1$, so $g_c=1/\lambda_+(J)$, approximately $1/(2\sigma)$ for Wigner disorder. Below this threshold, the dominant correlation decays as $C(t)\propto(g\lambda_+)^t=\exp(-t/\xi)$, with $\xi=-1/\log(g\lambda_+)\approx1/(1-g\lambda_+)$ near criticality. The controller maintains $g\widehat{\lambda}_+=1-\varepsilon$, where $\widehat{\lambda}_+$ is an online estimate and $\varepsilon\in(0,1)$ is the target stability margin.

Implementation notes

Integrate the controller into a tanh RNN, linear state-space model, or residual sequence block immediately before recurrent multiplication: $h_{t+1}=\phi(g_tWh_t+Ux_t+b)$. At every training step, estimate the largest singular value or largest real eigenvalue of the local Jacobian with 3--5 power iterations using a fresh Gaussian probe. For tanh, use $J_t=D_tW$, where $D_t=\operatorname{diag}(\phi'(g_tWh_t+Ux_t+b))$. Use the estimate $\widehat\lambda_{+,t}$ in the multiplicative update, with stop-gradient through the estimate. Pseudocode: initialize $g=0.9/\widehat\lambda_+$; unroll one minibatch; compute Jacobian-vector products; run power iteration; update $g\leftarrow g\exp[\alpha((1-\varepsilon)-g\widehat\lambda_+)]$; then backpropagate task loss and update network parameters. The paper supplies the spectral-edge mechanism and critical scaling; finite-width and nonlinear Jacobian edges must be estimated empirically. First test 256-unit tanh RNNs on copy memory and sequential MNIST, comparing standard orthogonal initialization, fixed spectral normalization, and the controller with $\varepsilon\in\{0.05,0.1,0.2,0.4\}$. Measure hidden autocorrelation, gradient norm versus sequence length, loss, and estimated edge. The quantitative prediction is a stability boundary near $g\widehat\lambda_+=1$: below one, correlation time follows $\xi\approx1/(1-g\widehat\lambda_+)$; above one, state or gradient variance grows rapidly. The measured critical gain should agree with $1/\widehat\lambda_+$ within approximately 20%, and the controller should hold the edge near $1-\varepsilon$.

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 8/10

Built a spectral-edge controller using power iteration on the local tanh Jacobian and verified the mechanism numerically. The toy sweep found the stability transition between effective edges 0.99 and 1.01, matched the predicted growth slope and correlation time to floating-point precision, and measured Wigner-edge/critical-gain scaling within 4.1%. The controller moved an initial effective edge of 2.0 to the target 0.9, but in the nonlinear recurrence it reduced gradient norm versus the fixed-gain baseline (0.00965 vs 0.219), so this is evidence for control and stability rather than a demonstrated performance win.

Agent confidence: 8/10

Baseline: Nonlinear fixed gain: mean effective edge 0.9374, gradient norm 0.21897, final state norm 0.6652
Idea: Online controller target 0.90: mean effective edge 0.9012, final gain 0.9507, gradient norm 0.009651, final state norm 0.5506; toy boundary crossing 0.99-1.01; Wigner edge and gain maximum relative errors 3.89% and 4.04%

Limitations: No copy-memory, sequential-MNIST, or trained task experiment was run; the nonlinear test used one small fixed-input tanh recurrence and a symmetric normalized matrix. The toy correlation law uses the exact dominant eigenmode, so its near-zero error validates the formula but does not establish finite-width nonlinear task benefits.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.26279) Sep 2, 2026

Beats the tuned baseline by 50.5% (p=0.0324, wins 7 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0031
Idea mean: 0.0016
Effect (Δ): -0.0016 (−50.5%; negative = idea better)
Wins: 7 / 8 paired seeds
p-value: 0.0324 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the required dynamics track was tested; vision, tabular, and sequence tracks were not run because the idea's structural match is stability/control dynamics. The controller uses a compact analytic local-Jacobian approximation for the manually unrolled GRU rather than a full exact Jacobian, and the benchmark used 12 training epochs and 400/200 samples rather than larger-scale training.

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 Worked ✓

Agent confidence: 8/10

Built the Spectral-Edge Criticality Controller as a local recurrent training intervention and evaluated it against a matched fixed-gain recurrent baseline on the dynamics track. Across 8 paired seeds, the controller reduced test MSE from 0.0031334 to 0.0015524, with delta_mean -0.0015810 and permutation p=0.0324, satisfying the benchmark's significant-win criterion. The trained-model mechanism signature confirmed the predicted effective edge target: observed 0.89979 versus predicted 0.9, relative error 0.024%.

Baseline: Dynamics rnn_small, best lr=0.006, 12 epochs: mean test MSE 0.0031333728547906503
Idea: Spectral-edge controller, best lr=0.006, target=0.9, alpha=0.25, 12 epochs: mean test MSE 0.0015523835318163037; paired delta -0.0015809893229743466; 7/8 wins; permutation p=0.0324; observed effective edge mean 0.899786445939283; mechanism confirmed=true

Benchmark result