Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Add auxiliary constraint nodes generated from linear combinations of existing constraints, creating a new message-passing graph while preserving the original feasible error set. Use a neural BP layer on the augmented graph so auxiliary nodes provide alternate paths around harmful cycles without changing the target constraints.
Add a learnable cyclic-coordinate mechanism to latent dynamics so that selected latent coordinates do not enter the Hamiltonian and their conjugate momenta become conserved. This provides an explicit dimensionality-discovery and invariance bias, encouraging the model to represent nuisance or symmetry directions compactly instead of spending independent dynamics capacity on them.
Replace a fully connected neural SDE drift with coordinate-wise functions that can read only the paths of graph parents. Learn soft edge gates and penalize violations of the paper's pathwise Lipschitz condition, so the model remains stable during long rollouts and supports explicit interventions on selected coordinates.
Insert a small continuous-time Markov latent module between a neural encoder and decoder, with input-dependent transition rates and a fixed library of graph topologies such as directed cycles, reversible chains, and branching motifs. The output is an observable of the stationary distribution, while a learned convex mixture over topology-specific response curves constrains the network to represent responses as combinations of interpretable nonequilibrium mechanisms.
Insert a data-fitted PCA bottleneck followed by a sparse multivariate Hermite polynomial head for a Gaussian-like latent representation. The head explicitly represents low-order and selected high-order interactions, while PCA controls high-dimensional input and output truncation error instead of forcing a generic MLP to learn these structures from scratch.
Replace an unconstrained recurrent matrix by a structured asymmetric circulant coupling whose Fourier modes have analytically known complex eigenvalues. A selected nonzero mode becomes a rotating attractor, providing a phase-coded recurrent state that can preserve information through oscillatory dynamics without requiring the optimizer to discover a stable spectral structure from scratch. A weak input projection and optional mode-selection loss can use the attractor as a nonlinear memory…
Replace dense self-attention by a sparse mask on a one-dimensional or ordered token geometry, retaining local neighbors and adding long-range edges with probability proportional to distance raised to \(-(1+\sigma)\). Use \(\sigma\approx0.8\text{--}0.85\) as the initial regime because the paper finds that this range supports delocalized, GOE-like connectivity despite sparse bonds. The resulting layer has linear or near-linear attention cost while maintaining long-range paths.
Build a positively homogeneous convex network by representing every intermediate unit as a compact polytope and composing units with Minkowski sums, convex-hull unions, and positive dilations. This gives an explicitly convex and monotone architecture whose geometric complexity can be controlled independently of the number of sampled linear pieces, potentially producing smaller ICNNs for structured convex functions.
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.
Partition tokens into a rectangular grid and use Latin-square labels to define several sparse attention heads. Each head attends only to tokens sharing a row, column, or Latin label, while orthogonality guarantees that every pair of labels occurs at most once, reducing systematic blind spots and repeated collisions. The resulting masks are deterministic, reusable across examples, and can be generated without learned routing scores.
Construct a weight-tied transformer loop in which the recurrent state receives a bounded diagonal carry plus a learned block increment, rather than applying a residual identity inside the learned increment. Parameterize the carry so every channel is strictly below one, allowing many recurrent iterations without the state explosion observed with an unconstrained carry.
Initialize a recurrent or state-space transition matrix with weak Wigner noise plus a shared cumulative-sum correlation structure. Increasing the correlation strength produces recurrent eigenmodes one at a time at analytically predicted BBP thresholds, yielding a controllable hierarchy of short- and long-memory modes. The matrix should then be globally rescaled or constrained so that all active modes remain inside the desired stability radius.
Replace the Euclidean hidden-state update of a recurrent or state-space neural network with a mixed manifold state containing a rotation component and Euclidean features. Propagate uncertainty with sigma points in tangent error coordinates, retract rotational perturbations through the exponential map, and compute the training loss from the predicted covariance. This avoids invalid rotations and captures second-order curvature effects that a first-order EKF-style recurrent cell misses at large…
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.
Construct a graph of cheap prediction agents or reasoning traces and use a sparse set of expensive verifier calls as graph anchors. Select the next verifier location by the exact reduction in a trace-inverse coherence objective per unit cost, rather than by uncertainty or random sampling. This creates a budgeted mixture-of-agents architecture that can spend computation where it most improves global consensus.
Build a neural PDE surrogate that predicts changes in equilibrium variables rather than changes in conservative state variables. The network receives the local state and geometry, predicts an equilibrium-coordinate increment, and subtracts the network output evaluated at a reference equilibrium, forcing the reference state to have exactly zero learned residual.
Replace interpolation of heterogeneous sensor streams by a phase-indexed recurrent or state-space network with period M, where M is the least common multiple of the sensor sampling periods. The network applies a distinct transition for each phase while using a fixed cyclic phase update, preserving timing structure and allowing missing observations to enter only when their phase-specific sensor is available.
Route tokens to experts using Laguerre cells defined by the minimum control energy needed to move a token embedding to each expert prototype, rather than by Euclidean distance or an unconstrained learned router logit. Per-expert dual weights deform the cells so that minibatch routing follows prescribed expert capacities, giving a geometrically interpretable alternative to auxiliary load-balancing losses.
Replace ordinary pairwise attention similarity by an affinity averaged over transformed keys or values. The resulting attention is invariant to the group action on either input and avoids requiring the network to learn identical attention patterns for every rotated or transformed copy.
Replace independent pairwise feature matching across augmented or multimodal views with jointly estimated soft permutation matrices constrained to agree through cycles. The paper's multi-view result suggests that independent copies can cross a correspondence-recovery threshold even when every individual pairwise matching is statistically non-informative. In a neural network, this can provide cleaner token, patch, object, or cell alignment targets and can be used either as a differentiable…
Replace an expensive proximal activation or implicit optimization layer with a Gaussian barycentric estimator computed from energy evaluations. The resulting map is smooth and has a provable cocoercivity guarantee when the energy is weakly convex, making it a stable alternative to unconstrained learned activations or iterative proximal solvers.
Measure how validation forecast error grows with prediction horizon and fit exponential and Mittag-Leffler models. When the Mittag-Leffler fit is decisively better, activate a fractional-memory SSM or long-memory residual branch and use its fitted effective order to set the branch's kernel decay and horizon-loss weights; otherwise retain a conventional recurrent or finite-memory branch.
Replace heuristic graph positional encodings with exact finite-abelian-group coordinates derived from edge-class increments and cycle constraints. Relative positions become group differences, allowing a graph transformer to share parameters across repeated generator displacements while retaining exact path consistency and compact cyclic coordinates.
Train a residual network on a coarse depth mesh, estimate a dual-weighted residual for every layer interval, and insert new layers at intervals with the largest estimated contribution to objective error. This replaces uniform depth expansion or expensive neural architecture search with targeted refinement driven by both forward-dynamics error and downstream loss sensitivity.