Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Construct a sparse fixed orthogonal mixer by repeatedly applying pi/4 rotations to randomly matched pairs of feature coordinates. Place this mixer before top-k feature pruning, sparse projection, or activation quantization so that information is spread across coordinates without using a dense random matrix.
Build each nonlinear correction in an inverse neural operator from explicit bilinear products of learned operator features, following the inverse Born expansion instead of using an unconstrained pointwise MLP. Use a square activation to implement multiplication exactly, and truncate the interaction order so the model has a controllable polynomial structure.
Replace the first several convolutional blocks of a small image model with a finite-depth convolution-modulus scattering stem built from a Parseval filter bank. Enforce exact energy accounting and use the paper's polynomial residual law to choose the smallest depth that captures the desired fraction of input energy, avoiding unstable or redundant deep scattering paths.
Replace dense token-to-token attention in a controlled ablation with a cyclic order-a recurrence whose coefficients are periodic across positions. The resulting mixer has linear cost in sequence length for fixed recurrence order and can enforce a fixed signed periodic boundary condition, providing a compact structured alternative to local attention or a lightweight state-space model.
Attach a finite mixture of zonotopes to each uncertain neural input or hidden state, and propagate every mixture component through affine layers and conservative nonlinear relaxations. When the number of components grows, merge components only with an enclosing zonotope and sum their probability masses, preserving a formal lower bound on the probability that the true activation lies in the represented set.
Represent hierarchical or tree-structured hidden states on words over d symbols and replace a dense mixing layer by a noncommutative Toeplitz operator composed of shared word shifts. Coefficients are reused at every tree location, so the parameter count depends on maximum interaction depth rather than the number of nodes; an optional spectral penalty controls the amplification profile of finite-depth truncations.
Replace independent top-k token selection by a quality-weighted determinantal subset objective based on the Hardy–Szegő kernel. Tokens with high learned quality are preferred, but geometrically redundant tokens have a small determinant contribution, encouraging diverse sets of routed experts, retrieved items, or attended context tokens.
Represent hidden features using a tensor-product polynomial-evaluation code instead of storing one value per feature. Corrupted coordinates can then be identified through violations of low-degree consistency and repaired before the next neural layer, targeting robustness to hardware faults, unreliable memory, malicious distributed workers, and adversarial activation corruption.
Replace an unconstrained multiplicative interaction between two nonnegative neural features by a lifted gate whose first and second moments satisfy the paper's semidefinite relaxation for the set F = {(x1,x2): x1,x2 >= 0, x1 x2 <= 1}. Insert the gate into an MLP, attention score, or MoE router to prevent explosive feature products while retaining a tractable convex feasible set.
Partition a neural state or feature vector into blocks and identify directed dependencies between blocks from one-step transition data. Use the inferred design structure matrix as a hard mask or soft gate on recurrent, state-space, graph, or mixture-of-experts couplings, replacing a dense unconstrained interaction matrix with a data-supported sparse graph.
Replace an unconstrained entrywise nonlinearity on a positive Gram or covariance matrix by a learned scalar function satisfying the paper's finite-order positivity-preserver conditions. The transformed matrix remains PSD for matrices of the target width n, allowing nonlinear Gram propagation without eigenvalue clipping or projection.
Replace or augment a dense attention similarity matrix with a Min-cone matrix generated by a monotone scalar sequence. The resulting matrix is positive semidefinite by construction, has only O(n) learned scalar parameters, and can be multiplied by values in O(n d) time using cumulative sums rather than forming an n-by-n matrix.
Replace the length-L learned convolution kernel in a causal sequence layer with K Laguerre basis functions, where K is much smaller than L and the basis parameter controls the decay time scale. The layer retains a long receptive field but learns only K coefficients, while FFT or a fixed state-space realization evaluates the resulting convolution efficiently. This is especially appropriate for audio, sensor streams, and long-context regression where the desired impulse response is smooth or…
Pack two real tensors into the real and imaginary components of one complex tensor, perform one complex transform, and recover both real spectra through conjugate symmetry. Combine this with implicit padding so separate padded real arrays and transform workspaces are not allocated.
Represent a 3D neural field using high-order polynomial coefficients attached to an adaptively refined tetrahedral mesh, with a small MLP predicting residual corrections from local coordinates. Refine only tetrahedra whose prediction, rendering, or PDE residual is large, and use globally ordered vertices so neighboring tetrahedra share identical face and edge coefficients without hanging-node constraint solves.
Replace a long sequence of graph propagation steps used to approximate a spectral graph filter with one block Krylov projection. Construct a basis from the input node features together with a small number of Gaussian probe vectors, evaluate the desired matrix function only on the resulting small projected matrix, and retain the output columns corresponding to the original features.
Add a distributed spectral positional encoding to a graph neural network, graph transformer, sparse-attention model, or MoE router by computing the dominant eigenvector of the current weighted adjacency matrix with a few warm-started power iterations. Unlike a Fiedler-vector feature, this encoding uses only local neighbor aggregation, is naturally nonnegative for nonnegative adjacency weights, and can be updated incrementally when the graph or edge weights change.
Augment a graph neural network with structural features computed from counts of small pattern homomorphisms whose pattern vertices are constrained to lie in selected vertex subsets. Unlike ordinary local aggregation, these features encode dense subgraph structure and can separate graphs or node sets that have identical low-order neighborhood statistics. Use a small learned bank of pattern graphs and sampled subset tuples so the method remains practical.
Replace a dense translation-invariant interaction matrix with a positive-definite Toeplitz kernel K_n(e^f) whose log-spectrum is parameterized by a small number of Fourier coefficients with 1/|k| decay. Use the paper's explicit quadratic term as a spectral-volume budget, allowing long-range structure while discouraging uncontrolled determinant growth and ill-conditioning. Subtracting this term from a log-determinant regularizer leaves a residual intended to capture higher-order deviations from…
Replace ordinary absolute positional embeddings with coordinates on a learned flat torus and use dual-lattice Fourier characters as positional features. Control the covariance of the coordinate fundamental domain so that the paper's inequality guarantees a lower bound on the smallest nonzero positional frequency, preventing the learned periodic coordinate system from developing arbitrarily weak or nearly constant modes.
Replace random or k-means initialization of a k-expert router with a moment-based range finder on a calibration batch of hidden states. Estimate a low-dimensional second-moment subspace, enlarge it using one-free-index third-Hermite contractions, and fit the router's expert centroids and weights only in this resulting subspace. The router can then operate on projected hidden states while retaining an optional small residual adapter.
Build neural computation graphs with explicitly phase-budgeted serial and parallel branches, treating serial compositions as SRG products and parallel residual branches as SRG sums. Allocate phase centers theta_i so that every loop or branch aggregate stays away from -1, enabling stability-aware architecture search and constructive control of branch gains.
Replace an unconstrained q-way polynomial or tensorized feature layer with separate decomposable and primitive interaction channels. The decomposable channel models interactions explainable as products of lower physical-weight feature blocks, while the primitive channel captures residual factors that cannot be represented by those products. This should reduce redundant high-order parameters and provide a controllable inductive bias for compositional or disentangled representations.
Replace a random cyclic filter bank or patch projection with the Weyl–Heisenberg orbit of one normalized learnable prototype. Regularize the prototype so that all nonzero shift and modulation correlations have a large and nearly equal magnitude, maximizing the smallest eigenvalue of the induced feature Gram matrix and preventing poorly observed feature directions.