Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Treat the binary outputs of the hyperplane head as a noisy channel and decode with reliability-weighted likelihood rather than unweighted Hamming distance. Estimate each bit's flip probability on validation data and give unreliable hyperplanes less influence, while retaining the logarithmic code-length scaling.
Replace a $K$-class softmax with $N$ binary hyperplane heads, where each class is represented by the signs of its projections onto fixed random directions. Train the embedding to reproduce these codewords and decode by nearest Hamming codeword. The paper's guarantee suggests that $N\approx 2\log_2 K+\log_2(1/\delta)$ can separate all class centers with high probability in sufficiently high dimension, giving a concrete width rule rather than choosing the number of binary heads heuristically.
Initialize a coordinate-network feature bank with the leading eigenfunctions of a bandlimited concentration operator instead of random Fourier features. For a desired spatial region E, these features maximize the fraction of their L2 energy inside E among all functions with frequency support in Omega, giving a principled basis for localized signals.
Use a transport map \(Q_\theta\) from a fixed latent reference distribution to a data distribution, but expose only its locally averaged version \(\bar Q_{\theta,\sigma}(z)=\mathbb E_{u\sim K_\sigma(\cdot-z)}Q_\theta(u)\). Latent-space mollification integrates the pole-type influence singularity instead of allowing one training sample near \(Q_\theta(z)\) to dominate the quantile feature or its gradient.
Replace a single Laplacian or fixed diffusion regularizer in a CNN with a finite positive mixture of fractional Laplacians at several orders. The resulting module separately controls short-range smoothing and long-range spatial coupling, while positivity preserves a dissipative energy and avoids the unstable behavior of arbitrary signed mixtures.
Construct a scalar feature or critic for oscillator-based neural dynamics that is invariant under the transformations imposed by free harmonic motion and elastic collisions. For finite-size rods, the module should represent only quantities compatible with common oscillator-phase rotations and momentum permutations, preventing a learned world model from inventing coordinate-dependent pseudo-conserved quantities that disappear after collisions.
Replace the ordinary combinatorial Hodge propagation in a simplicial neural network with a geometry-induced weighted Hodge Laplacian built from Euclidean simplex volumes. The operator preserves the harmonic/topological subspace while changing the positive spectrum according to the shape and scale of the simplices, allowing message passing to distinguish geometrically meaningful cells that have identical incidence patterns.
Replace an unsigned two-hop cyclic mixer by the paper's alternating signed circulant. The sign pattern preserves one-step and two-step interactions while reducing the exact spectral radius from 4 to 2√2, allowing a larger raw mixing coefficient under the same operator-norm stability constraint.
Replace or augment a residual neural layer with a Fourier-domain scale-selective flow containing a learned second-order term and a fourth-order stabilizer. The block permits controlled low-frequency amplification, as required by the KS infrared mechanism, while damping high-frequency feature noise and preventing unbounded spectral growth.
Insert a learnable semigroup layer that evolves features according to a positive operator combining frequency damping and spatially varying confinement. Unlike isotropic Gaussian smoothing, the layer can damp selected frequencies differently along different axes and can suppress activations in learned spatial regions, while the positive-semigroup construction prevents amplification.
Replace the unsigned adjacency used by a deep message-passing network with a signing selected from an affine family that makes designated short even cycles unbalanced. Search this family for a small even-power trace, which acts as a proxy for a smaller spectral radius and suppresses explosive long-range propagation. The signing can be fixed before training, so the method adds no per-example inference cost.
Replace a dense node-to-landmark graph-attention or message-passing relation by a dictionary of distinct landmark-neighborhood signatures. Nodes sharing the same signature reuse the same structural landmark aggregate, while their individual hidden states are still passed through the output MLP, preserving node-specific predictions. On bounded-treewidth graphs the number of distinct signatures is provably linear in the number k of landmarks, with an explicit dependence on treewidth.
Construct attention groups as hyperedges of a linear r-uniform hypergraph: every pair of tokens is allowed to share at most one group, while each token participates in approximately the same number of groups. Apply local attention inside each group and aggregate the outputs across groups. The construction inherits the paper's sharp capacity bound and prevents both redundant pair interactions and high-degree token hubs.
Represent a nonnegative attention or routing score matrix by two nonnegative vectors, X = uv^T, and learn only entries on a sparse bipartite graph of important query-key or token-expert interactions. Complete the remaining entries multiplicatively and monitor cycle residuals as a certificate of whether the sparse representation is compatible with rank one. Use local ratio violations to trigger additional edges or relax the rank-one approximation only where needed.
Replace ordinary k-means-style prototype assignment with a distance-decay capture layer whose scale varies across samples, tokens, or classes. Train with a cooperative concave surrogate over prototype centers and anneal toward hard nearest-prototype assignment; this explicitly preserves useful gradients for multiple nearby prototypes while retaining sparse facility-like behavior at inference.
Replace ordinary row-degree or symmetric normalization in a directed graph neural network with a nonlinear Dyson scaling. For a nonnegative directed adjacency matrix A, solve a positive vector equation and propagate with B = D A D, where D is the diagonal matrix of the solution. The resulting operator has row sums strictly below one, giving an explicit bound against exploding directed message propagation while retaining asymmetric edge information.
Replace a learned embedding for each integer with a compositional embedding of its variable-length Gray codeword. Add an auxiliary constraint that numerically adjacent values have nearby representations, while preserving the ordinary task loss so that the model can learn when numerical adjacency matters.
Replace a dense channel-mixing matrix by a sparse complex generalised weighing matrix W with exactly w nonzero entries in every row and column, then use U=W divided by square root of w as a norm-preserving mixer. Restricting to k=2 gives a real matrix with entries in {+1,-1}; k=4 supports signed phase rotations. The exact isometry should preserve signal and gradient norms while reducing channel-mixing cost.
Replace a standard softmax MoE router with a thermodynamic router whose expert occupations maximize entropy subject to a prescribed total routing mass and mean routing energy. At high temperature, traffic is distributed across many experts; as temperature decreases or the energy budget tightens, traffic undergoes a predictable condensation transition in which excess load moves to the lowest-energy expert or expert group. This supplies an explicit control knob for adaptive specialization instead…
Use an anisotropic singular relative-position kernel in attention or graph message passing, with its exponent constrained by the paper's local integrability threshold. The module can represent sharper directional interactions than an RBF while providing an explicit certificate that its spatial gradient belongs to a chosen L^p space.
Replace a parameter-heavy recurrent transition, or use this as a fallback, with a two-parameter nearest-neighbor successor blend in latent space. Given a query latent state, retrieve the closest state from an in-context trajectory and combine the query, the retrieved state, and its observed successor; this gives a zero-shot dynamical forecast with almost no trainable transition parameters.
Replace one independently learned vector per graph edge with a truncated spectral expansion on the line graph. The model learns coefficients for low-frequency edge modes and reconstructs edge features before message passing, reducing parameters while imposing an inductive bias that incident edges should have correlated behavior.
Replace mean-only readout from a noisy recurrent or Langevin reservoir by concatenating empirical first, second, and fourth raw moments of each hidden coordinate. The second and fourth moments retain input-dependent width and tail information generated by nonlinear confinement, while multiple independently initialized reservoirs can be concatenated before the final linear classifier to preserve complementary features.
Insert a differentiable implicit layer that maps boundary features to an interior latent field by solving a discrete sinh-Gordon equation. The paper's second-order convergence result motivates using a symmetric five-point discretization and a damped Newton solve rather than asking a neural network to learn the entire interior field directly.