Unverified 2026

Criticality-controlled mutation schedule

Implementation & benchmark of arXiv:2609.03123 — Death by mutants: unusual multicritical dynamics in a two-species model for absorbing state transitions

Usefulness5/10

Difficulty4/10

Novelty8/10

Training Stability Generalization Sample efficiency Training dynamics Scheduler Regularization Rnn Control theory Dynamical systems Bifurcations Statistical mechanics

Source paper: Death by mutants: unusual multicritical dynamics in a two-species model for absorbing state transitions arXiv:2609.03123 ⓘ · analyzed Sep 4, 2026

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 Not peer-reviewed.

Idea description

Use the two-species phase diagram as a training controller: gradually increase one-way mutation $\lambda$ and mutant suppression $\mu$ while monitoring whether feature activity approaches the absorbing boundary. This creates a measurable curriculum from independent learning to controlled competition instead of turning on strong destructive interactions at initialization.

Formulas

$$\partial_t\rho_A=\cdots-\mu\rho_A\rho_B,\qquad \partial_t\rho_B=\cdots+\lambda\rho_A,$$

$$M_s(t)=\frac{1}{N}\sum_{i=1}^{N}\rho_s(i,t),\qquad \widehat g_s(t)=\frac{\log(M_s(t+W)+\varepsilon)-\log(M_s(t)+\varepsilon)}{W},\quad s\in\{A,B\},$$

$$\lambda_{q+1}=\operatorname{clip}\left(\lambda_q-\eta_\lambda\widehat g_B(q),0,\lambda_{\max}\right),$$

$$\mu_{q+1}=\operatorname{clip}\left(\mu_q+\eta_\mu\widehat g_A(q),0,\mu_{\max}\right).$$

Mathematical statement

The decoupled limit is obtained with $\lambda=\mu=0$, while coupling adds B production $\lambda\rho_A$ and A suppression $-\mu\rho_A\rho_B$. Define empirical activities $M_A$ and $M_B$ as spatial and batch averages of the nonnegative hidden fields. Estimate their logarithmic growth rates over a window W. The controller increases mutation when B is decaying too quickly and increases suppression when A is growing too quickly, targeting near-zero growth without allowing either activity to vanish. The paper's multicritical point is near $a_1=b_1=0$, and its distinct A/B scaling means the controller must monitor both species separately rather than use one global feature norm.

Implementation notes

Integration point: apply this controller to the coupling coefficients of the nonreciprocal feature block during optimization. Keep a1, b1, a2, and b2 fixed initially, and update lambda and mu only every Q optimizer steps using batch-averaged activities. Pseudocode is: for optimizer_step: y,A,B=net(x,lambda,mu); loss=task_loss(y,target); loss.backward(); optimizer.step(); if step%Q==0: MA=mean(A.detach()); MB=mean(B.detach()); gA=(log(MA+eps)-log(MA_prev+eps))/window; gB=(log(MB+eps)-log(MB_prev+eps))/window; lambda=clip(lambda-eta_lambda*gB,0,lmax); mu=clip(mu+eta_mu*gA,0,mmax); MA_prev=MA; MB_prev=MB. The paper supplies the coupled and decoupled limits and the multicritical interpretation; the growth-rate feedback law is an empirical control adaptation. Compute activities from actual hidden tensors, smooth them with an exponential moving average, and freeze the controller during validation and testing. First cheap experiment: use a two-layer recurrent classifier on sequential MNIST, comparing zero coupling, fixed strong coupling, random coupling, and the adaptive schedule. Sweep initial a1 and b1 around zero. The predicted signature is a stable nonzero-activity plateau with both estimated growth rates near zero under feedback, while fixed coupling should cross into extinction or runaway activity as mu and lambda increase. Identify the empirical absorbing boundary by plotting extinction probability against coupling strength; the adaptive schedule should remain within one activity standard deviation of this boundary. Near the boundary, A traces should show a statistically detectable power-times-log correction relative to pure-power fits if the paper's critical mechanism survives neural discretization.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.