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 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 independent edge-type logits in a relational graph neural network with a mean-field fixed-point router derived from a colored ERGM. Each edge's color distribution is influenced by its own relation bias and by the expected number of rainbow triangles it forms with neighboring edges, allowing the model to learn coordinated multilayer structures.
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…
Replace ordinary Wasserstein or arithmetic pooling of distribution-valued features with a barycenter whose individual quantile displacements are Huberized. Small changes between input distributions remain averaged quadratically, while a corrupted token, expert, graph neighborhood, or augmentation cannot move the pooled distribution arbitrarily far. The module is especially cheap for one-dimensional distributions represented by fixed quantile vectors.
Build a kernel aggregation layer whose output is a tangent vector field on the unit sphere and whose surface divergence is identically zero by construction. For each source point, use a matrix kernel obtained by applying a surface-rotated gradient in the query variable to a scalar zonal kernel; this is a differential-form version of the paper's matrix-valued construction. The layer can replace attention or message passing when the target dynamics are incompressible, such as spherical fluid…
Use the relaxed QFT tensor-network topology as a trainable norm-preserving mixer inside a neural block, replacing a dense token-mixing matrix or an expensive global convolution. The network learns data-adapted global interactions while retaining structured O(N log^2 N) application and an exact cheap inverse, making it suitable for image tokens, long sequences, or reversible residual blocks.
Add a causal memory branch whose lag-response function is represented by a Bernstein polynomial with coefficients constrained to produce a nonnegative, decreasing, convex kernel. The branch aggregates past hidden states using this kernel, giving the model a learnable long-memory profile while preventing oscillatory, negative, or increasing historical influence.
Replace unconstrained token-mixing logits by a symmetric zero-row-sum response matrix generated from positive conductances on a small auxiliary electrical network. The resulting mixer has conservation and positivity structure, while circular minors have a prescribed sign pattern associated with positive grove measurements. This is especially suitable for graph neural networks and attention variants that need stable global diffusion rather than arbitrary dense affinities.
Replace an unconstrained attention or MoE routing matrix with a row-stochastic matrix that preserves a learned Gibbs-like distribution over groups. Anneal its temperature so that the router converges to a block-upper-triangular operator: mixing remains flexible within equal-energy groups, while cross-group traffic becomes directional rather than oscillating or collapsing. This creates a hierarchy-aware inductive bias and can reduce the number of active expert or attention connections.
Replace the linear phase progression in a positional encoding or rotary attention mechanism with a deterministic quadratic phase. The resulting position signal is generated by an irrational rotation with linearly changing increments, and the paper proves that its infinite diffraction measure is purely absolutely continuous, suggesting disorder-like spectral coverage without random sampling.
Attach each token or graph node a learned scalar charge q_i and add a fractional stable kernel K_ij = exp(-tau D |q_i-q_j|^alpha) to the interaction mechanism. Constrain 0 < alpha <= 2, the exact range in which the kernel is positive semidefinite for arbitrary finite real charge sets, and optionally make tau layer-dependent to obtain multiscale interactions. This provides a principled alternative to unconstrained learned distance biases and can be used either as an attention-logit bias or as a…
Construct a neural mixing layer only from Brauer generators for the orthogonal group: identity, pairwise swaps, and pairwise contractions with the Euclidean metric. This gives an exactly O(2)-equivariant alternative to unconstrained tensor mixing, with trainable coefficients but fixed symmetry-preserving basis maps.
Replace flat expert logits with scores generated by Gaussian increments on a binary routing tree. A leaf receives the sum of increments on its root-to-leaf path, so sibling experts have correlated logits and the router can learn nested coarse-to-fine specialization; an inverse-temperature schedule controls the transition from exploratory diffuse routing to sparse routing.
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.
Parameterize a sequence attention graph by a finite graph motif on positions [r], then lift that motif to arbitrary sequence lengths using strictly increasing embeddings. Instead of learning an independent positional mask for every context length, the model learns a small collection of order-preserving edge types whose induced graph is valid at every length and can be extended without retraining.
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.
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 freely learned finite impulse-response mixing kernel with a matrix polynomial whose roots are constrained to the unit circle. The resulting block-Toeplitz operator has an explicitly positive semidefinite spectral construction, while increasing the polynomial degree gives a systematic capacity knob for approximating matrix-valued frequency responses.
Replace an unconstrained matrix nonlinearity on small symmetric feature blocks with the isotropic spectral lift of a permutation-equivariant monotone map on eigenvalues. The layer remains orthogonally equivariant, while the paper's equivalence transfers a scalar inner-product monotonicity certificate from eigenvalue space to the full matrix space.
Replace independent Gaussian attention noise or unconstrained token routing with a directed-polymer path distribution over positions and layers. The router aggregates exponentially many monotone paths through temporally correlated random edge scores, producing heavy-tailed but spatially coherent routing and preventing attention from collapsing onto a single token. The paper's t^{2/3} wandering and t^{1/3} free-energy fluctuations become measurable diagnostics and tunable targets rather than…
Replace an unconstrained learnable distance-bias function in a graph neural network or distance-aware attention layer by a Bernstein approximation of a positive-definite circular kernel. The resulting kernel is a degree-n polynomial in normalized distance while preserving positive semidefiniteness of every finite Gram matrix on the circle, preventing training from producing an invalid covariance-like similarity structure.
Replace raw pairwise attention or graph-edge scores by exact U-centered residuals, removing additive effects attributable to either endpoint. The resulting scores represent interaction beyond independent source and destination biases and satisfy zero row sums, preventing a few high-degree or high-activation tokens from dominating relational aggregation.
Compute one-dimensional persistent homology and minimum-norm harmonic representatives, then use their absolute edge coefficients as topology-aware saliency in a graph transformer. Add the saliency to attention logits or use it as a soft regularizer so the model preferentially propagates information along edges that are essential to persistent cycles.