Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Use a symplectic Hamiltonian update as a recurrent or state-space neural block, preserving a learned modified energy across many layers or time steps. This targets residual and recurrent architectures where ordinary Euler updates accumulate drift during long rollouts.
Transform local neural residuals into the Ripa model's characteristic coordinates before spatial aggregation, apply a mode-dependent gate based on neighboring characteristic jumps, and transform back. This lets the model damp oscillatory acoustic or equilibrium-mode corrections near discontinuities without globally smoothing every feature.
Parameterize a learned feature-space operator as accretive but not necessarily symmetric, then apply its fractional power through a finite positive mixture of shifted resolvents. This provides a matrix-function layer that can represent directional and rotational interactions while avoiding unstable eigendecomposition of nonnormal matrices.
Represent a hidden state as complex-valued points on a two-dimensional lattice and replace unconstrained local updates by the exact harmonic-quadrilateral completion rule from discrete conformal geometry. Given three corners of a plaquette, compute the fourth corner by a Mobius-rational formula enforcing cross-ratio minus one, then use a learned readout or forcing term for task-specific predictions. The layer supplies a hard geometric inductive bias and a directly measurable local constraint…
In a partially observed reinforcement-learning or model-based control agent, expose the state-estimator innovation to the action head through a dedicated residual feedback branch. The policy produces a nominal action from the estimated latent state, while a learned innovation-compensation branch corrects actions when observations disagree with predicted latent dynamics. This explicitly separates nominal policy behavior from estimation-induced corrections and should help during fast transients…
Insert a differentiable spatial canonicalization module before a neural dynamics model. It estimates a smooth invertible coordinate transformation that places each input field in a common gauge relative to a reference template, predicts the next state in that gauge, and maps predictions back to the original coordinates. The module should reduce the need for the dynamics network to relearn identical laws under many smooth spatial reparameterizations.
Replace a dense recurrent transition matrix with a periodic CMV-style product of alternating local 2x2 unitary cores. The transition is exactly norm-preserving, has O(n) trainable parameters under periodic tying, and can be applied through local factor operations rather than stored as an n-by-n matrix. Use turnover refactorization when changing the ordering or boundary connection of cores, enabling a compact cyclic unitary state-space layer.
Replace a generic recurrent transition with two coupled unitary transitions that share one block column and differ by a sign on the other block column. Each transition preserves hidden-state norm exactly, while the structured difference gives a controlled two-path recurrent architecture for long-context modeling.
Extract a small set of stable exponential modes from an observed neural sequence and use them to initialize a diagonal or block-diagonal state-space model. Hankel-pencil eigenvalues propose the modes, while persistence across shifts and contour margins reject modes caused by noise or a short-lived background.
Train an unconstrained branch and a geometry-aware branch in parallel, then learn how much to trust the analytic branch. This preserves the benefit of explicit geometry on correctly specified tasks while allowing the model to ignore a misleading or irrelevant prior.
Replace an unconstrained feature vector entering a rational or resolvent-like neural operator by a polynomial feature whose first nonzero Taylor coefficient lies in a pole-safe subspace. For a pole of order m, the simplest guaranteed construction is psi(z)=(z-beta)^m v, which makes Q(z)psi(z) bounded even when Q(z) diverges. For lower-order cancellation, solve linear constraints among Taylor coefficients of psi so that all negative Laurent powers vanish.
Insert a scalar flux-correction-style limiter after a neural operator predicts a conservative state or residual. Interpolate between a known-admissible baseline state and the learned high-order candidate, choosing the largest coefficient that satisfies a geometric family of linear inequalities encoding positive density, positive pressure, and subluminal velocity. This retains as much of the neural prediction as possible instead of independently clipping physical variables.
Parameterize a relative-position or lag-decay function as a finite positive mixture of exponentials instead of learning arbitrary attention bias values. The resulting kernel is completely monotone on positive distances, so it is nonnegative, decreasing, and has alternating derivative signs; the mixture provides several learned memory scales without allowing oscillatory or unstable long-range biases.
Construct a latent layer whose node states are small positive-definite matrices and whose local updates follow a weighted cluster exchange relation rather than an unconstrained affine map. The update is reversible when the old state is retained, while noncommuting matrix products preserve relational structure that scalar cluster variables cannot represent.
Store a convex object as a direction-indexed vertex tuple and implement composition of objects through componentwise Minkowski addition and nonnegative scaling. This creates a structured residual or compositional layer where convexification is nonexpansive, making perturbation amplification controllable and avoiding repeated generic geometric optimization.
Replace one deterministic residual update with a short cyclic composition of learned vector fields evaluated for randomized, short run times. Because finite compositions of noncommuting flows generate directional-derivative and Lie-bracket terms, changing the cycle order gives the network an explicit, low-cost way to learn drift directions that are unavailable from the individual vector fields alone.
Constrain a channel-mixing layer to be a product of nonnegative bidiagonal matrices, rather than an unconstrained dense matrix. The resulting totally nonnegative operator is predicted not to increase sign oscillations in ordered channel features, potentially reducing high-frequency feature noise and making deep stacks more stable.
Replace a dense token- or channel-mixing matrix with a product of local braid generators acting on adjacent coordinates. Each generator is an exactly invertible 2-by-2 transformation, while the braid and far-commutativity identities give multiple equivalent factorizations of the same global operator. This creates a sparse, reversible mixer with O(kn) cost for a braid word of length k, rather than O(n^2) cost for a dense matrix.
Insert a fixed reversible lattice shear into a residual network so successive blocks follow a structured monodromy orbit rather than using unrelated learned transformations. Apply the transformation to a small learned subspace of hidden channels while leaving the remaining channels unchanged. This creates deterministic phase-dependent feature mixing with no additional trainable parameters.
Replace a deterministic mixture-of-experts residual block with K population-indexed stochastic expert states coupled through a graphon matrix. The layer uses a shared drift and expert-dependent diffusion, while an empirical convex-order penalty makes later representations more dispersed than a reference representation without permitting a mean shift.
Use the paper's eventual path-length bounds to constrain an order-invariant routing graph to a constant-hop communication budget. A learned sparse attention or graph-neural-network layer can explicitly route information through at most three admissible hops, while a more conservative auxiliary route permits at most five minimal-path hops, preventing increasingly long and unstable dependency chains as sequence length grows.
Replace an ordinary elementwise interaction between two feature matrices by a noncommutative functional-calculus layer \(\varphi(A,B)\), where \(A\) and \(B\) are Hermitian channel operators that need not commute. Add a soft penalty on \([A,B]=AB-BA\), and use a Besov-smooth parameterization of \(\varphi\) so that perturbations are controlled in Schatten \(p\)-norm for \(p\leq2\). This creates a principled matrix interaction module that can remain stable when feature operators or graph…
Represent a continuous-time neural dynamical system as a symbolic Markov chain over regions together with a positive learned roof function giving the time spent in each region. Weight local reconstruction and prediction errors by the predicted vector-field speed, following the paper's scaled Hölder coding relation, so that the model does not over-penalize arbitrarily small coordinate errors near equilibria. This produces a hybrid latent model with discrete long-range structure and continuous…
Insert a fixed or partially learnable equivariant change-of-basis module into a spherical or SO(3)-equivariant network. At each angular frequency \(\ell\), the module maps the line selected by the line-bundle quantization to the line selected by the Grauert-tube quantization, allowing the network to represent both holomorphic/base-local and geodesic-flow-adapted features without breaking rotation equivariance.