Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Replace a recurrent token-mixing operation with a finite-state carrier scan over binary or quantized token features. The local transition table is constrained to conserve a scalar token weight and to be bijective, making the mixer reversible, constant-memory, and less prone to activation drift than a generic recurrent layer.
Replace a dense weight matrix by a cross approximation built from selected rows and columns rather than by a conventional truncated SVD. Periodically refresh the selected indices using residual leverage scores, warm-starting from the previous factorization so that the compressed layer can track weight changes during fine-tuning.
Build an input-conditioned unitary transformation as an ordered product of exponentials of anti-Hermitian matrices, with each factor controlled by a univariate function of one input coordinate or one learned scalar projection. This replaces a dense multivariate matrix-valued controller with separable scalar nonlinearities while preserving exact unitarity at every forward pass.
Replace a flat MoE or graph-pooling assignment with recursive partitions selected by interaction evidence after removing each item’s expected degree effect. Tokens, nodes, or examples that are frequently active for purely popularity-related reasons should not automatically form an expert or cluster. Recursion stops when a candidate split has nonpositive degree-corrected evidence, producing an adaptive hierarchy rather than a fixed number of equally sized groups.
Replace a dense channel or token-mixing matrix with a product of positive bidiagonal factors, so information propagates through a controlled sequence of local couplings rather than arbitrary signed interactions. Initialize the factors from the paper's barycentric-subdivision factorization, then learn positive diagonal and off-diagonal parameters; the resulting map is structured, parameter-efficient, and constrained to remain totally positive.
Replace ordinary one-token-to-one-expert or one-token-to-one-attention routing with a local latent subset router: a pooled observation can be explained by a compatible subset of tokens. Pairwise compatibility scores assign probability to subsets, and each token receives the marginal probability that it belongs to the selected subset. This should help when tokens represent overlapping objects, occluded entities, or multiple features that should be processed jointly.
Replace a flat stochastic MoE router by a two-level continuous-time routing model: experts within a group communicate rapidly, while transitions between groups occur slowly. Use the effective class-level stationary distribution as a soft load-balancing prior, reducing routing oscillation while preserving expert specialization.
Parameterize a relative-position or lag-decay function as a finite positive mixture of exponentials instead of learning arbitrary attention bias values. The resulting kernel is completely monotone on positive distances, so it is nonnegative, decreasing, and has alternating derivative signs; the mixture provides several learned memory scales without allowing oscillatory or unstable long-range biases.
Replace an independently generated discrete latent sequence or redundancy stream with a symmetric two-state Markov source whose transition probability is tuned or learned. Train the downstream transformer to reconstruct the semantic target after random insertions, while using the paper's insertion-capacity expression to select the latent rate and redundancy budget. The representation should preserve information under timing drift, repeated tokens, and inserted distractors better than iid token…
Replace dense attention between grid-arranged tokens by a deterministic block-sparse pattern generated from modular permutations. In each block, connect row token i to column token p(i)=2i modulo B; because i, i-p(i), and i+p(i) are injective modulo B when gcd(B,6)=1, the pattern avoids repeated horizontal, vertical, and diagonal projections. Use shifted permutations across heads to increase receptive-field coverage while retaining structured sparsity.
Replace random Fourier features or a dense sinusoidal positional encoding with a compact bank whose frequencies are the continued-fraction denominators of an irrational number. Inverse-frequency amplitudes provide multiscale structure with a controlled sub-Lipschitz regularity profile, while lacunarity reduces the number of frequencies needed to represent oscillatory structure.
Use the attention probability distribution over an ordered context to choose contiguous token groups whose pooled attention masses have entropy as close as possible to a prescribed upper budget R. Replace the corresponding key/value vectors by one weighted representative per group, preserving token order and reducing the KV-cache length from n to m. Unlike unconstrained token merging, the entropy constraint gives a direct control knob over how concentrated or diffuse the retained attention…
Replace dense token-to-token attention by a learned binary relation generated from a small number of hierarchical predicates, while rejecting masks that contain a fixed K_{t,t} biclique. The paper's incidence bound predicts near-linear active edges for these structured relations, giving sparse attention with a measurable worst-case complexity target rather than relying only on average sparsity.
Replace ordinary masked mean pooling with a Fourier-compressed quadrature operator for arbitrary two-dimensional or three-dimensional domains. The geometry is preprocessed once into reusable grid weights, allowing every channel and every training example using the same domain to be pooled without boundary-area bias.
Estimate the temporal spectrum of each sequence channel using a locally private procedure, then apply a regularized inverse-square-root spectral filter before the sequence enters attention or an SSM. The filter removes predictable low-frequency or narrow-band redundancy while avoiding unstable amplification at frequencies where the private estimate is small.
Replace unconstrained combinations of several attention or adapter operations with a brace-style composition that inserts each operation into a distinct ordered interval of a base sequence. The resulting computation preserves the order of host and inserted operations and forbids crossing dependencies, producing hierarchical attention patterns with an explicit structural bias.
Replace raw braid-generator sequences by sequences of positive simple Garside factors obtained from the left-greedy normal form. Because powers of \(\Delta\) lie in the Hilden subgroup, they can be removed while preserving the relevant double-coset presentation, reducing non-uniqueness and often shortening the sequence. Feed the resulting factor tokens to a Transformer or sequence classifier, and train it to be invariant to inserted removable \(\Delta\)-powers.
Represent candidate two-dimensional attention windows as dyadic rectangles and penalize local regions where many deeply embedded windows overlap. Use complementary horizontal and vertical depth exponents rather than independently penalizing one coordinate. The resulting router should reduce pathological concentration of sparse attention computation while preserving access to multiscale context.
Replace the usual hand-designed expert-load penalty with a heterogeneous survival penalty derived from a susceptibility distribution. Each expert receives an availability factor q_e=G(A_e), where A_e is its cumulative recent routing pressure and G_e is a learned or fixed mixture of exponentials; highly used experts are suppressed smoothly, while heterogeneous experts can have different resistance to pressure. The mixture produces adaptive curvature and long-tailed penalties that may reduce…
Replace a conventional scalar activation by a geometrically indexed family of affine pieces whose slope changes with the logarithmic magnitude of the input. The same two endpoint parameters are reused across all scales, giving a compact, explicitly scale-aware activation that can represent different responses for exponentially separated activation magnitudes.
Replace an unconstrained Cartesian-product router over heterogeneous branches with a router whose joint expert or state assignments obey a finite-group conservation rule. Branch i emits a distribution over labels in its own subgroup H_i of a common finite abelian group G; only tuples whose group sum is zero are retained. This gives an exact, differentiable structural prior for modular arithmetic, multi-relational graphs, multi-view fusion, or any setting where latent labels compose by a…
Replace a dense or ad hoc sparse attention pattern with a circulant mask generated by the paper's carry word c(m,k). Every query attends to exactly m of k relative positions, and the selected positions are prefix-balanced, avoiding the large gaps and collisions produced by random sparsification. Use several Farey-related slopes across heads or layers to combine local, medium-range, and long-range coverage.
Build a constrained autoregressive model whose initial logits are generated from a translation-invariant MPS associated with a local zero-mode construction. The MPS supplies a structured valid distribution before a Transformer residual is added, so the model starts on the constraint manifold instead of learning validity through a penalty.
Partition activations into dyadic magnitude bands and allocate sparse connectivity separately to heavy and diffuse coordinates. Protect high-magnitude coordinates with more reliable connections while using randomized flat connectivity for the many small coordinates, keeping the total number of nonzeros fixed.