Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Partition activations into dyadic magnitude bands and allocate sparse connectivity separately to heavy and diffuse coordinates. Protect high-magnitude coordinates with more reliable connections while using randomized flat connectivity for the many small coordinates, keeping the total number of nonzeros fixed.
Approximate a graph's adjacency by a learned abelian Cayley host and use one shared message-passing operator for every edge in the same inverse-pair generator class. Keep only the unexplained original edges as a residual branch, so the layer interpolates between a parameter-efficient group convolution and ordinary graph message passing.
Before message passing, repeatedly detect a pair of vertices with nested open neighborhoods and fold away the dominated vertex while preserving its information in the surviving vertex's feature state. The graph reduction is justified by homotopy invariance of the independence complex, while the feature merge prevents task-relevant attributes from being lost. Add a topology-aware ablation comparing this exact fold against random node pooling and standard learned pooling.
Replace a channelwise wavelet or strided-convolution front end with vector-valued wavelet filters that deliberately pair different scalar wavelets across channels. The resulting subbands retain compact-support multiscale structure and can be recombined exactly, while a small learned 1x1 mixing layer operates on the cross-channel coefficients instead of learning a full expensive convolution at every scale.
Replace an unconstrained low-rank adapter or similarity projection with a learned subspace carrying a prescribed signed metric. The module learns an orthonormal basis U for a k=p+q dimensional subspace, forces the compressed form U^*I_{m,n}U to have p positive and q negative eigenvalues, and uses the resulting pseudo-inner product for signed attention or retrieval scores.
Replace random or learned routing hashes for tokens arranged on a d by d grid with a fixed family of mutually orthogonal anti-Latin squares. Each channel assigns exactly d of the d squared tokens to every bucket, while any two channels jointly distinguish every grid position. The resulting router has deterministic load balance and multi-view positional diversity.
Add a task-aware error-protection code to a binary or low-cardinality latent representation. The encoder remains systematic, preserving the original latent coordinates, but appends repeated or parity coordinates computed from a linear task map so that latent states with different task values are separated by at least a chosen Hamming distance. Redundancy is allocated according to the rank of the task map rather than the full latent dimension.
Replace a dense learned polynomial-feature transform with a d-orthogonal recurrence whose production matrix is constrained to a (d+2)-banded lower-Hessenberg form. The layer generates successive features using only local recurrence coefficients, giving O(dN) arithmetic and O(dN) parameters for N basis functions instead of O(N^2) dense mixing.
Parameterize a learned token metric as a nonnegative sum of sparse integral rank-one projections with unimodular support, rather than learning an unconstrained dense positive-semidefinite matrix. Graph-incidence covectors give an immediately implementable support family, while nonnegative coefficients guarantee positive semidefiniteness by construction.
Build a sparse recurrent graph-neural layer on a path-by-path, path-by-cycle, or cycle-by-cycle latent lattice using a skew-zero-forcing seed set and its forcing order as a causal update schedule. Only the currently forced target node is activated at each step, so a small number of anchor states can propagate through the complete lattice while retaining local connectivity and periodic-boundary structure. The exact seed-count formulas predict the minimum number of anchors required by the graph…
Build a fixed multiscale router that maps 2D coordinates to 3D voxel coordinates using the paper's X-shaped self-similar refinement. Use the router to run a 3D feature field or volumetric token mixer over a 2D-organized tensor, while retaining a mathematically controlled locality bound instead of an arbitrary flattening permutation. The first target is a 3D neural field or small voxel classifier where the router replaces either a dense 3D feature table or a naive raster-order token layout.
Construct a learnable spectral pooling block as an erosion followed by its adjoint dilation, making the resulting opening idempotent, increasing, and anti-extensive. The block can suppress frequencies outside a learned passband while guaranteeing that applying it twice does not continue changing the representation, which is useful in multi-stage CNN pyramids and U-Net skip paths.
Interpret successive neural representations as an RG flow and constrain coarse-graining layers to remove unstable or redundant information monotonically. The paper reports monotonic decrease of an effective central charge along measurement-induced RG flows; a neural analogue can use a measurable information-complexity proxy and reject compression steps that increase it while preserving task-relevant information.
Replace dense attention on tree-structured inputs with stochastic attention neighborhoods formed by metric balls of sampled radii. Use the paper's exact trimming rule to ensure that every sampled cover remains valid while its total radius budget is bounded, then average predictions over several independent covers during training. This creates sparse, globally covering attention masks with an explicit locality-versus-coverage control.
Replace an unrestricted GRU or attention-based history encoder with a fixed companion-form shift register driven by the current action and observation, followed by a learned nonlinear policy. The register stores a structured finite history, while a learned matrix or MLP readout maps that history to a control-relevant latent state. This should provide a cheaper and more interpretable memory mechanism for partially observed environments, especially when the relevant dynamics are approximately…
Build a binary hierarchy over tokens by recursively splitting each active block with a beta-splitting rule, then perform dense attention only inside small leaf blocks and communicate between leaves through learned summaries at internal nodes. The beta parameter controls how balanced the partition is, while the paper's maximum-depth asymptotic supplies a principled depth budget and a way to detect pathological trees.
Construct multiplicative neural gates directly on encoded tensors so that operands are multiplied coordinatewise without decoding between every operation. Polynomial evaluation makes this operation algebraically consistent with multiplication, allowing redundant gated MLPs or bilinear layers to retain fault tolerance while reducing the frequency of expensive correction steps.
Compress a module whose output changes with a scalar condition such as diffusion time, temperature, or compute budget by representing its response in a low-rank basis generated by resolvent-like functions. Distinct spectral modes produce rational factors \((1-\tau\lambda_k)^{-1}\), allowing a small number of learned components to approximate a large hypernetwork or condition-dependent parameter table.
Constrain a sparse attention graph to split into a k-degenerate backbone and a residual graph with maximum degree at most k-1. Orient the backbone according to a degeneracy order so that each token receives or emits at most k structured interactions in the relevant direction, while the residual edges form a bounded-degree correction layer. This replaces arbitrary sparse attention with a topology that is easier to schedule and whose worst-case edge and local-degree costs are explicit.
Replace a complex latent vector x in C^d by squared magnitudes of m learned complex linear projections. Set m equal to 2d: the paper proves that m less than or equal to 2d minus 1 cannot generically preserve the latent up to global phase, whereas m equal to 2d is generically sufficient, giving a principled minimal width for a phase-invariant neural bottleneck.
For a neural model whose outputs lie on a d-dimensional analytic family in a very high-dimensional space, replace the full output vector by 2d+1 or a modestly oversampled number of fixed Gaussian scalar measurements. The paper's theorem predicts almost-sure injectivity in the noiseless setting, so an inverse network or decoder can recover the same latent instance without processing the full observation. Because the theorem does not provide a noise-stability constant, use M=4d+8 or M=8d in the…
Construct a sparse token-mixing architecture by interleaving learned per-token transformations with fixed perfect-shuffle and cyclic-pile permutations. For n not a power of k, the generated permutation group is 2-transitive, so sufficiently rich sequences of generator words can expose every ordered token pair without constructing a dense N by N attention matrix.
Replace an unrestricted additive recurrent or fast-weight memory with a sign-selectable update: for each incoming update vector, choose between adding and subtracting it so that a smooth compact potential of the memory state is minimized. This is appropriate when the memory representation has sign symmetry, such as signed random features or a learned linear sketch; it is not a drop-in replacement for ordinary gradient updates where the sign carries semantic information.
Replace an unconstrained geometric multiscale codebook by features generated from a finite digit set and a Pisot scale factor. The contracting algebraic-conjugate directions should suppress near-collisions between representations at different scales, producing a discretely separated hierarchy that can be used for embeddings, recurrent memory, or quantized transformer states.