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.
Construct a recurrent or state-space layer as a skew product: an expanding bounded feature coordinate drives a linearly contracting hidden state. Constrain the hidden transition matrix A to have spectral radius below one, and monitor the predicted transition ell times the absolute determinant of A equals one: below it, hidden trajectories should occupy a thin or fractal set, while above it they should have substantially higher-dimensional state coverage without losing local contraction.
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.
Replace an unconstrained recurrent block with two coupled modules: a contractive perceptual estimator and an input-to-state-stable cognitive state transition. Spectral normalization and a controlled Euler residual step enforce a quantitative gain condition, preventing hidden-state explosion while retaining long memory when the contraction factor is chosen close to one.
Construct a generative or recurrent neural architecture with several contractive or mildly expanding branches, and explicitly control the geometric complexity of its invariant set using the sub-additive singular-value pressure of branch-Jacobian products. Instead of regularizing only the operator norm, the model can preserve anisotropic directions while targeting a desired attractor dimension, potentially improving coverage of structured data without uncontrolled folding or collapse.
Attach a symbolic sparse head to a neural encoder instead of using a dense final MLP. The head evaluates a library of learnable power-law and interaction terms on nonnegative learned features, jointly optimizes linear coefficients and exponents, and removes inactive terms with coefficient sparsity. This should provide a compact model with better relative-error behavior on positive targets spanning several orders of magnitude.
Replace unconstrained output coordinates with a neural parameterization whose outputs are valid monotone profiles by construction, analogous to representing a Young diagram through nonnegative ordered row increments. Train the network against an explicit energy or negative log-probability while preserving the feasible geometry, rather than relying on penalties that permit invalid intermediate states.
Represent an attention or routing state as a symmetric projector or fixed-spectrum positive semidefinite matrix and refine it using the paper's double-bracket flow instead of unconstrained gradient steps. The update rotates the state toward a task-derived Hermitian cost matrix while preserving its eigenvalues, so rank, trace, and spectral diversity remain fixed by construction.
Periodically project a rectangular neural-network weight matrix onto an approximately orthonormal-column matrix using LU-preconditioned CholeskyQR rather than ordinary QR or a polar iteration. Pivoted LU handles badly scaled and nearly dependent columns, while Householder orthogonalization of the LU factor produces a triangular preconditioner that makes the subsequent Cholesky step safer in fp16 or bfloat16.