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 ordinary topology-sensitive message passing with scalar-gated aggregation followed by an explicit correction that aligns local node states with a graph-wide consensus component. The correction should make node embeddings less sensitive to line or edge removals while preserving local information needed for prediction. This is suitable for graph neural networks and graph-based world models exposed to changing graph sizes or sparsity patterns.
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 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.
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…
Replace full-state quantized write-back in a deep low-bit residual stack with quantized increment error feedback. The residual branch quantizes the proposed increment after adding the previous carry, while the carry stores the exact discrepancy; this makes the total error telescope instead of accumulating approximately once per layer.
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.
Replace a deep feed-forward block by the fixed point z=phi(Wz+Vx+b), with the recurrent weight W constrained so that the fixed point is unique for every input. The same condition makes forward fixed-point iteration stable and makes implicit differentiation well-conditioned, allowing depth-independent memory usage while providing a measurable spectral failure boundary.
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.
Replace an unconstrained token mixer or feed-forward residual map with a feature-conditioned rational operator whose transfer matrix is positive real on the positive orthant. Apply it through a resolvent, rather than an additive residual, so that the accretivity certificate yields a nonexpansive implicit update and suppresses activation explosions.
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 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.
Construct deep or recurrent networks whose layer weights are correlated across depth with a prescribed power-law covariance, rather than either fully tying or fully independently sampling layers. The paper predicts two usable design boundaries: \(\gamma=1/2\) for divergence of correlation-induced fourth moments and \(\gamma=1\) for loss of summable-correlation flatness.
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.
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.
Replace a fixed-size learned latent for an array-valued complex tensor with a variable-length list of continuous rank-one spectral atoms. An encoder predicts candidate receive direction, transmit direction, residual off-grid offsets, and complex gains; the decoder reconstructs the tensor analytically from the array-response formula, so changing the antenna dimensions does not require changing the decoder weights.
Replace an unconstrained trajectory or density network with a stack of RealNVP-style triangular coupling layers whose inverse and log-volume change are analytic. Condition the coupling subnetworks on the task prompt and time, so the same invertible module represents task-specific population states while providing an exactly computable density and score surrogate.
Replace pointwise sequence reconstruction with reconstruction of overlapping past and future Hankel windows in a shared latent manifold. A first encoder compresses the delay-coordinate trajectory, while a second decoder or predictor reconstructs the future block from the latent state; training therefore penalizes representations that fit observations but do not preserve dynamical evolution.
Replace a conventional whitening transform with a constrained whitening layer that minimizes cross-channel covariance while requiring every output channel to remain correlated with its designated input channel by at least a threshold \(\rho_{\min}\). The layer exploits the orthogonal freedom in whitening to find a rotation that preserves channel identity instead of arbitrarily mixing features. It can be inserted before an MLP, convolution, or attention projection and compared directly against…
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…