Unverified 2026

Nonreciprocal mutant feature fields

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

Usefulness6/10

Difficulty5/10

Novelty7/10

Architecture Stability Accuracy Generalization Cnn Rnn Training dynamics Regularization Statistical mechanics Dynamical systems Pde Stochastic processes

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 Not peer-reviewed.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

Replace a single residual feature field with two nonnegative streams: a primary representation A and a mutant or corrective representation B. A produces B through a one-way source term, while B suppresses A, creating an explicitly tunable competition mechanism that can prevent feature collapse and encourage complementary representations.

Formulas

$$\partial_t\rho_A=D_A\nabla^2\rho_A+a_1\rho_A-a_2\rho_A^2-\mu\rho_A\rho_B+\xi_A\sqrt{\rho_A},$$

$$\partial_t\rho_B=D_B\nabla^2\rho_B+b_1\rho_B-b_2\rho_B^2+\lambda\rho_A+\xi_B\sqrt{\rho_B},$$

$$\rho_A^{k+1}=\operatorname{softplus}\!\left(\rho_A^k+\Delta t[D_A\nabla^2\rho_A^k+a_1\rho_A^k-a_2(\rho_A^k)^2-\mu\rho_A^k\rho_B^k]+\sqrt{\Delta t}\,\xi_A\eta_A^k\sqrt{\rho_A^k+\varepsilon}\right),$$

$$\rho_B^{k+1}=\operatorname{softplus}\!\left(\rho_B^k+\Delta t[D_B\nabla^2\rho_B^k+b_1\rho_B^k-b_2(\rho_B^k)^2+\lambda\rho_A^k]+\sqrt{\Delta t}\,\xi_B\eta_B^k\sqrt{\rho_B^k+\varepsilon}\right).$$

Mathematical statement

Use the paper's coupled reaction-diffusion dynamics with nonnegative feature densities $\rho_A(x,t)$ and $\rho_B(x,t)$. Here $D_A,D_B\geq0$ are spatial mixing strengths, $a_1,b_1$ are linear growth rates, $a_2,b_2>0$ are self-saturation coefficients, $\mu\geq0$ suppresses A through B, $\lambda\geq0$ produces B from A, and $\xi_A,\xi_B$ are noise amplitudes. The essential nonreciprocal structure is the one-way source $\lambda\rho_A$ in the B equation and the competitive term $-\mu\rho_A\rho_B$ in the A equation. Positive quadratic saturation prevents unbounded pointwise growth. Near $a_1=b_1=0$, both populations approach the absorbing state; the paper predicts ordinary directed-percolation-like scaling for B and logarithmically modified scaling for A below its upper critical dimension.

Implementation notes

Integration point: insert this as a recurrent residual block between ordinary convolutional or transformer feature projections. Split an input feature tensor into two groups, map both through softplus to obtain nonnegative fields A and B, and evolve them for K small Euler-Maruyama steps. Use a depthwise 3x3 Laplacian convolution for the diffusion terms; parameterize a2 and b2 with softplus so they remain positive, and parameterize lambda and mu as nonnegative scalars or channelwise vectors. Pseudocode is: A=softplus(A0); B=softplus(B0); for k in range(K): etaA=normal_like(A); etaB=normal_like(B); A=softplus(A+dt*(DA*lap(A)+a1*A-a2*A*A-mu*A*B)+sqrt(dt)*xiA*etaA*sqrt(A+eps)); B=softplus(B+dt*(DB*lap(B)+b1*B-b2*B*B+lam*A)+sqrt(dt)*xiB*etaB*sqrt(B+eps)); y=projection(concat(A,B)). The reaction equations, asymmetric coupling, and square-root demographic noise come from the paper; the Laplacian stencil, softplus positivity map, finite-step solver, and output projection are engineering adaptations. Begin with zero noise to isolate deterministic behavior, then add noise during training. First experiment: compare this block with a standard residual block and a symmetric two-stream block in a matched-parameter ConvNet on CIFAR-10. Sweep a1 and b1 around zero while measuring mean activity, extinction probability, Jacobian spectral radius, and accuracy. The predicted signature is an active-to-absorbing transition near a1 approximately equal to zero and b1 approximately equal to zero. Near that point, A activity should fit C*t^(-alpha_A)*(log t)^p_A better than a pure power law, whereas B should be closer to pure-power decay. Reject the mechanism if no transition is visible or if the logarithmic fit does not improve held-out residual error by at least 10 percent.

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.