Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Use mutually orthogonal Latin labels as deterministic expert assignments for token batches. Each routing family is individually balanced, and pairs of families avoid repeated co-assignment patterns, enabling multiple routing rounds or auxiliary experts without the severe load collisions caused by independent random hashing.
Replace activation-magnitude-based adaptive computation halting with a criterion based on the actual recurrent update and a local stability margin. The loop halts when the state change is small relative to state scale for several consecutive steps, avoiding pathological decisions when LayerNorm-driven dynamics cause the activation norm to collapse.
Replace independent top-k MoE routing with a submodular polyhedral allocation over experts. A learned set function assigns a marginal gain to each additional expert allocation, so the router exhibits diminishing returns and can enforce global capacity constraints rather than making unrelated per-token choices. The allocation is obtained by sorting marginal gains, giving a fast greedy router with piecewise-linear routing regions.
Decompose tensor-valued hidden states into invariant symmetric, alternating, and higher Young-symmetry channels before applying learned maps. This removes redundant tensor coordinates and prevents a neural layer from mixing incompatible representation types, yielding smaller equivariant modules with a cleaner inductive bias.
Construct a reusable ReLU trunk that emits approximate univariate powers or Legendre-polynomial features for each input coordinate, then combine them with a linear or low-rank polynomial head. This gives a compact explicit basis for smooth functions and can replace a large generic MLP in low-dimensional scientific regression or serve as a frozen or partially trainable front-end.
For smooth coordinate-based regression, replace a width-heavy MLP with a deliberately narrow but deeper ReLU network and choose depth and width using the paper's analytic-function approximation law. The hypothesis is that, at fixed parameter count, increasing depth gives a larger reduction in approximation error than increasing width when the target is close to analytic.
Replace an unconstrained Fourier-feature block in an implicit neural representation or coordinate MLP with a sparsity-aware layer whose output gain is normalized according to the distance outside the training interval. The normalization uses the paper's endpoint law, preventing a small in-domain Fourier representation from producing arbitrarily large outputs just beyond the observed coordinate range.
Replace independent coordinate rounding of a fixed-sum vector with nearest-point quantization in the projected integer lattice A_n^*. The quantized vector preserves the zero-sum constraint exactly, while the globally optimal rounding correction accounts for the aggregate residual induced by projection.
Replace a polynomial graph filter or repeated matrix multiplications in a graph neural network with a small rational filter evaluated at several shifts. Treat the incoming feature matrix as a block of probes rather than processing scalar probe vectors independently, allowing one set of shifted solves to expose multiple spectral directions simultaneously. The expected gain is higher approximation quality at the same number of operator applications, especially when the target filter has sharp or…
Replace a single smooth inverse predictor near detected ambiguity boundaries with multiple prediction branches and a soft gate. The gate is trained to preserve distinct decompositions rather than forcing the network to interpolate through a thin high-curvature transition layer, while a Jacobian or curvature penalty identifies unresolved ambiguity regions.
Add a coordinate-aware long-range aggregation branch whose singular low-frequency component is explicitly centered before it is mixed into token representations. The centering acts as a neural counterterm: constant or slowly varying value fields cannot accumulate an activation contribution that grows with context size, while local and higher-frequency interactions remain available through an ordinary attention residual branch.
Turn a recurrent or state-space memory into a constrained hereditary state: the latent state remains in a learned convex domain, and only input motion that reaches the boundary changes the plastic component. This creates a nonexpansive, rate-independent memory that should suppress unstable state growth and make the representation depend on meaningful cumulative changes rather than arbitrary update frequency.
Augment each token or graph node with a periodic latent position x_i and phase θ_i, then evolve these variables before attention or message passing. Tokens with similar phase attract in x, while tokens with similar position synchronize in θ, producing self-organized groups without an externally specified clustering objective. The coupling strengths J and K provide interpretable controls for aggregation and synchronization, and their sweep should expose the paper's four collective regimes and…
Replace independent uncertainty heads in a branching neural network with a structured variational posterior whose non-root node distributions condition on jointly sampled latent states of all parents. This allows collider evidence to explain away upstream uncertainty: evidence at a child can alter the posterior over several parent branches instead of leaving their uncertainties artificially independent. The approach can be implemented as a stochastic DAG network and trained with an evidence…
Replace or augment geometric attention on spatial or point-cloud tokens with a positive fractional kernel containing the paper's inverse-square origin factor. This gives tokens near a designated singular center a controlled increase in receptive-field influence while preserving a scale-invariant distance decay, which may help models represent cusp-like fields, radial singularities, and multiscale spatial interactions.
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.
Replace a dense learnable Fourier multiplier with a low-parameter multiplier concentrated near the common zero set of two polynomial constraint symbols. A linear constraint together with a cubic constraint can produce straight or curved frequency loci, allowing the network to represent directional long-range structure while using far fewer spectral parameters than a full 3D frequency grid.
For risk-sensitive or recursive objectives, add a separate network that predicts the conditional certainty equivalent of the next-state continuation value, rather than forcing the value network to approximate a nested nonlinear expectation directly. Train the value, policy, and certainty-equivalent heads with Bellman and first-order residuals jointly.
Replace a fixed Fourier or spectral resolution in a neural operator or sequence model with a data-adaptive spectral cutoff. Keep only modes whose estimated signal energy exceeds the noise-amplification and discretization floor implied by the available number of trajectories and samples per trajectory. This should reduce overfitting to high-frequency sensor noise and preserve accuracy when the same model is deployed at different sampling resolutions.
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.
Add an online low-rank reconstruction layer before a recommender or ranking MLP. It estimates a latent action-feature subspace from masked candidate vectors, freezes that subspace for an epoch, reconstructs each partially observed candidate in the latent coordinates, and feeds only those coordinates to the predictor. The method is most promising when the ambient candidate dimension d is large but the effective rank m is small and missingness is not too severe.
Add a slow latent two-state gate to a recurrent, state-space, or world-model network so that separate experts represent two qualitatively different dynamical regimes. Train the gate using the paper's two-state population and fluctuation mechanism rather than allowing an unconstrained softmax to average incompatible regimes. The model should allocate extra capacity near the gate's susceptibility peak, where regime uncertainty and forecast variance are predicted to be largest.
Replace a large dense layer whose input and output dimensions factor into multiple modes by a TT-matrix whose parameters are stored as a chain of small cores. Periodically apply TT-SVD rounding to remove weak singular directions and keep the representation within a prescribed approximation error. This transfers the paper's central computational principle—perform tensor-product contractions directly in compressed form—to neural network layers.