Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Replace dense attention over structured object tokens with attention over role-filler tensor-product representations. A learned query specifies both a role and a filler, retrieves objects matching that binding, extracts a target role, and rebinds the extracted filler into an output object.
Replace the usual sum of pairwise modality similarities with a higher-order score based on the coordinatewise Hadamard product of all normalized modality embeddings. For modalities indexed by i=1,...,m, score a tuple using s(x_1,...,x_m)=\omega^\top(\bar g_1(x_1)\odot\cdots\odot\bar g_m(x_m)), where \omega is learned and \odot is coordinatewise multiplication. This adds explicit m-way interactions without concatenating raw features or introducing a joint encoder.
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.
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 the usual random or elementwise-positive linear-attention feature map with a rank-one positive-semidefinite feature map derived from query and key vectors. For normalized inputs, the resulting kernel is the squared inner product, which is nonnegative and gives a geometrically structured interference pattern that is better suited to associative recall than an arbitrary low-rank feature map.
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.
Give every possible pairwise relation its own recurrent latent state, stored in a dictionary keyed by stable node identities, instead of discarding the state whenever the edge is absent from the current graph. At each timestep, active edges retrieve their old state, update it with the current pair features and neighborhood messages, and write it back after message passing. This should improve predictions in dynamic graphs with intermittent interactions and reduce the burden on node states to…
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…
Replace ordinary row-softmax attention with a doubly stochastic Sinkhorn attention plan W, and periodically recover a gauge-fixed pairwise cost from W using the exact double-centering identity. Use this recovered cost to initialize or regularize a structured attention score, making the attention geometry identifiable despite arbitrary query and key row and column offsets.
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.
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.
Add a weighted reflection symmetry to an attention or graph-propagation matrix instead of requiring ordinary permutation equivariance. For paired positions or graph nodes related by an involution, penalize the failure of the propagation operator to commute with the weighted reflection; this makes all geometric multi-step propagations symmetry-compatible. The method is suitable for data with mirror, reversal, paired-agent, or left/right structure where the two sides have unequal importance…
Insert a linear Johnson–Lindenstrauss bottleneck around a set of jointly processed representations, choosing its width from the sharp finite-set dimension bound rather than from the model's nominal hidden size. The projection should preserve pairwise distances between tokens, patches, or retrieved items, allowing a downstream attention or MLP block to operate at lower width while retaining the geometry relevant to similarity computations.
Replace a wide collection of interchangeable near-zero branches with a module whose output is explicitly a quadratic form in the branch-weight Gram matrix. The module preserves the paper's leading-order behavior while making the relevant collective variable explicit and allowing low-rank parameterizations.
Parameterize an entropic OT cost only in directions that can change the transport plan, removing row-plus-column potential directions that are invisible because of OT gauge invariance. Whiten the remaining feature coordinates using their empirical covariance, producing an OT layer whose identifiable parameters have substantially more uniform sensitivity.
Replace multiplicative query-key attention scores with an affinity based on the l1 distance between first-spike latency vectors. For each query token and key token, small latency differences produce large affinity and distant timings decay exponentially, yielding a locality-sensitive attention pattern naturally compatible with leaky spiking neurons.
Freeze a randomly initialized single-layer transformer and use a constructed soft prompt to make its attention weights equal Gaussian-kernel weights over support examples. The resulting model performs Nadaraya-Watson regression in one forward pass, so task adaptation stores prompt tokens rather than modifying network weights. Prompt length becomes the number of kernel centers, while hidden dimension and prompt norm determine whether the required logits can be represented accurately.
Replace an n-by-n attention or token-mixing matrix with two nonnegative rank-r factors having row-simplex constraints and a shared latent column marginal. The induced matrix is exactly doubly stochastic at every accepted update, while applying it to values uses two thin matrix multiplications and never constructs the dense attention matrix.
Use predicted covariance reduction as a differentiable gate for selecting tokens, views, sensors, or retrieved demonstrations. The gate favors inputs with high expected information gain while accounting for acquisition cost, turning attention and data collection into active observability optimization.
Replace full agent-to-agent state transmission with coefficients in a learned dominant Koopman mode basis. Agents communicate only the leading spectral coordinates that explain slowly decaying collective behavior, while retaining a certificate based on the spectral gap and subdominant eigenvalue to decide whether the compressed representation is safe.