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.
Replace ordinary token merging or graph pooling with a learned block map whose output preserves information about a remote target conditioned on the surrounding coarse representation. The paper's majority-spin counterexample gives a concrete failure mode: two microscopic configurations mapped to the same pooled token can imply different predictions for distant variables.
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.
Constrain a student policy to transform its action in the same way that the input state is transformed, while constraining its value estimate to remain unchanged. During distillation, augment every teacher-student pair with several symmetry-transformed copies and penalize disagreement after transforming the student action back to the original frame.
Initialize each row of a neural weight matrix as a stationary correlated Gaussian process instead of using independent entries, but constrain its correlation tail to remain on the finite-fourth-moment side of the transition. This creates controllable structured spectra while avoiding the heavy-edge regime predicted for correlations slower than \(t^{-1/2}\).
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…
Replace a single residual stream or unconstrained hyper-connection with S parallel feature streams whose cross-stream mixing matrix is doubly stochastic. Parameterize the matrix with Sinkhorn normalization so every layer preserves total stream mass while still learning adaptive information routing. This is a low-overhead alternative to dense cross-stream attention and should reduce stream explosion, collapse, and sensitivity to depth.
Represent local feature channels as a smooth directional signal and aggregate them with a partition-of-unity family of learnable spherical atoms instead of hard angular bins. Use the atom Gram matrix to whiten the descriptor and add a projected-energy loss, so the network is rewarded for retaining information in the directional subspace rather than merely producing large correlated channel responses.
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.
Replace fixed PCA-rank selection in a hidden layer with a renormalization-group-inspired gate over covariance eigenvalue bands. The gate retains modes whose effective quartic interaction remains unstable or strongly scale-dependent, while pruning bands that flow toward the Gaussian noise fixed point. Unlike top-eigenvalue truncation, this is designed for extensive-rank signal distributed throughout the bulk spectrum.
Train a cheap shared multi-task probe briefly, extract one semantic embedding per task, and use density-based clustering to determine which tasks should share a neural trunk. After clustering, replace the globally shared trunk by one trunk per discovered cluster, with task heads remaining separate; this preserves cooperation among related tasks while isolating destructive task interactions.
Replace deterministic LoRA importance scores with one-sided tests of whether each rank-one update has population contribution at least a user-selected threshold. Maintain empirical contribution samples during fine-tuning, estimate their uncertainty, and prune the components with the weakest statistical evidence while respecting the target rank budget. The method should avoid deleting components merely because their latest minibatch gradient was small.
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.
Build a multi-expert or multi-task layer whose feature channels are divided into a globally shared subspace and expert-private subspaces. Matrix-weighted message passing couples experts only through selected feature directions, while the nullspace preserves specialization; the graph-cut condition provides a concrete test that the shared channels can propagate across all experts rather than becoming disconnected islands.
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.
Insert a projection-space mixer that combines several fixed or learned directions using reciprocal correlations with the current feature, then normalize the result. The exact construction has a universal beta law for its squared input-output cosine, so it can create controlled angular diversity while remaining deterministic and independent of the chosen direction dictionary.
Replace an unconstrained hidden-to-hidden interaction in an MLP or transformer feed-forward block by two gauge-related branches. Split channels with an orthogonal involution Θ, constrain the learned interaction K to anticommute with Θ, and use opposite signs of K in paired branches. This creates a testable inductive bias in which the learned interaction only transfers information between the two channel subspaces.
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.
Replace an MoE router's single softmax distribution with a normalized coordinate-wise product of several simplex-valued routing factors. The product preserves positivity and normalization but, as depth grows, concentrates mass on a small subset of experts, creating a mathematically controlled heavy-tailed routing prior rather than relying only on an auxiliary load-balancing loss.
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…