Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Replace Cox-de Boor evaluation of each cubic B-spline edge activation with its fixed truncated-power expansion. Normalize each scalar edge input to a bounded knot coordinate, evaluate the five shifted cubic positive-part terms in parallel, and contract them with the learned spline coefficients inside one fused kernel.
Insert a constraint-reduction layer between a policy network and its executed action. The policy proposes an action, while the layer retains only geometrically extreme collision and obstacle constraints and verifies that every discarded halfspace is implied by the retained ones through nonnegative conic multipliers. The reduced projection or quadratic program is therefore equivalent to the full tightened safety filter whenever certification succeeds, but uses substantially fewer constraints.
Construct a residual network from independently attachable modules, but permit only a number of modules whose aggregate feedback gain lies inside a delay-dependent admissible interval. Estimate deployed end-to-end latency and each module's local Jacobian gain, then reject or bypass additional modules when the predicted delayed-loop stability boundary is crossed. This turns variable-width or depth scaling into a falsifiable control problem rather than an empirical choice.
Replace black-box differentiation through an embedded LP decision with an analytic Jacobian computed from the LP’s active basis. A neural policy emits LP coefficients or right-hand sides; the LP returns the decision, while the backward pass uses the basis inverse and dual sensitivity, avoiding solver unrolling and finite-difference noise.
Replace unrolled autodiff through an ordered block-implicit neural layer with a custom reverse sweep that solves one small transposed local system per forward block update. The backward computes the exact gradient of the executed finite-depth solver while avoiding a global Jacobian and retaining only compact block information.
Replace a square dense projection in a Transformer or MLP with a trainable recursive butterfly matrix. The layer preserves multiscale channel interactions while constraining every complementary row-column block to rank at most k, reducing parameters and enabling recursive structured matrix-vector products. Unlike an arbitrary sparse layer, the construction has an explicit recursive factorization and a quasi-optimal approximation guarantee among matrices with the same butterfly rank.
Replace a dense neural-network weight tensor with a graph tensor network whose physical modes and internal edge ranks are specified by a sparse rank-adjacency matrix. Unlike tensor-train or hierarchical Tucker layers, the graph can contain selected cycles and skip connections between tensor modes, allowing the factorization topology to match correlations in the weight tensor. Fit the layer with GTN-SVD at a prescribed tolerance and compare accuracy, parameter count, and tensor-contraction…
Partition graph nodes into backward-equivalent classes and run message passing on the K-node quotient graph instead of the original N-node graph. If every node in a class receives the same aggregate message from every source class and shares the same local update map, class-constant node representations remain class-constant at every layer, making the quotient computation exactly equivalent to the full GNN on that invariant subspace.
Implement causal linear attention in chunks and combine chunk summaries with an associative scan instead of carrying the recurrent state through all chunks sequentially. This preserves the exact causal computation while reducing inter-chunk dependency depth from the number of chunks to its logarithm, enabling substantially more GPU parallelism for long-context training and prefill.
Replace recurrence or nearest-neighbour analogue lookup with a learned delay-coordinate observer that continuously corrects a latent state using the current observation. Constrain the observer's closed-loop Jacobian or linear state matrix to have spectral radius below one, so prediction error contracts geometrically and required burn-in grows logarithmically with target accuracy.
Replace a dense directed attention matrix by a collection of K learned source-to-hub-to-target interactions. Each hub corresponds to a directed biclique, allowing many source tokens to communicate with many target tokens using O(NK) rather than O(N^2) pair interactions. The construction preserves asymmetric information flow and can be initialized from a graph cover of high-attention edges.
Replace an autoregressive rollout of a learned dynamical model with a branch-trunk factorization that predicts all future steps simultaneously. The branch network encodes the future action sequence, while the trunk network encodes the current state and query coordinates; their inner products produce the complete horizon. This removes repeated state updates during inference and gives a compact differentiable model for planning.
For a recurrent or graph neural network with known local connectivity, estimate each node's local Jacobian row using only graph neighbors rather than all hidden coordinates. Use the resulting sparse Jacobian both to compute a contraction certificate and to regularize training toward dynamically local interactions, reducing estimator variance and the number of samples required for reliable stability decisions.
Compress each hidden layer by retaining directions that are simultaneously reachable from the observed input distribution and observable at the network output. Unlike PCA or SVD, the retained subspace is weighted by downstream task sensitivity, so high-variance but output-irrelevant directions can be removed while low-variance predictive directions are preserved.
Use a spherical-design codebook and the paper's polar slack factorization to create a nonnegative geometric interaction bias for attention or expert routing. The resulting kernel is generated by a rank-one term and a rank-at-most-d term, and entries close to zero can define a structured sparse mask instead of relying only on learned top-k selection.
Replace dense or single-dilation sparse attention with two sequential sparse attention stages whose offsets form co-prime arithmetic progressions. The first stage mixes tokens separated by multiples of M2, the second by multiples of M1; their composition reaches virtual offsets mM2+nM1, providing many structured long-range interactions from only M1+M2-1 physical offset families. Use causal masking and residual connections so the module can replace a standard transformer attention block without…
When a model, dataset, or parameterization has an involutive symmetry, transform gradients and curvature into symmetry sectors before applying adaptive updates. The Hessian and optimizer can then be handled as independent even and odd blocks, preventing curvature from one sector from forcing an unnecessarily conservative learning rate in the other.
Replace a quadratic pairwise attention or graph aggregation kernel with a compact, translation-invariant indefinite kernel approximated by signed random Fourier features. The feature map preserves the kernel's negative spectral mass through a diagonal sign matrix, so the resulting linear-time aggregation can represent similarities that ordinary positive-definite random features cannot.
Represent a large linear map acting on a Cartesian 3D grid and multiple physical channels as a TT-matrix, while retaining separate TT blocks for channel couplings that have different semantics. Apply the layer by sequential contractions with TT cores rather than materializing a dense matrix or a full 3D convolution kernel. Rank truncation provides an explicit accuracy-versus-memory knob and can be applied after optimizer updates.
Replace dense graph self-attention with two parallel branches: exact softmax attention only over graph neighbors and a global linear-attention branch that summarizes all nodes through feature-space statistics. A learned node-wise gate interpolates between the branches, allowing locally structured nodes to use sparse attention while retaining a global-information path.
Construct a cheap graph or token-mixing operator by partitioning nodes into k blocks using the bottom nonconstant eigenvectors of P squared, then replacing dense pairwise mixing with conditional averaging inside each block followed by one baseline propagation step. Unlike ordinary spectral clustering, the bottom modes target partitions where block labels are rapidly destroyed by P, producing an aggressively mixing representation layer rather than a community-preserving pooling layer. The…
Replace an arbitrary graph-attention mask with a fractional edge mask lying in the intersection of the spanning-tree polytope and twice the matching polytope. The mask represents a distribution over connected spanning trees while imposing expected degree at most two at every vertex, after which sampled trees can be used for sparse message passing.
Replace full PSD self-attention with a pivoted Cholesky/Nyström approximation whose landmarks are sampled from the unexplained diagonal mass. Tokens with large residual self-similarity are more likely to become landmarks, so the rank budget is spent on difficult regions rather than uniformly selected tokens.
Build a label-free router for a finite library of neural operators by estimating one shared physical target response from an anchor prediction and using it to rank every candidate through inner products with candidate differences. The method avoids running a full residual-based diagnostic independently for every expert and can be used either to select the best expert or to form a corrected weighted combination.