Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Add a learnable orthogonal rotation to a hidden representation and train it to make every channel projection have a small ψ2/L2 ratio. Unlike variance normalization, this explicitly suppresses directions with unusually heavy empirical tails while preserving the total quadratic energy of the representation.
Add a nonlocal logarithmic-Laplacian penalty to intermediate spatial feature maps or ordered token embeddings. Unlike a standard graph or image Laplacian, the kernel uses scale-free weights proportional to |z|^{-n} and includes a local compensation term, allowing multiscale feature smoothing without simply forcing nearby features to become identical.
Build a differentiable assignment layer whose rows represent tokens and whose columns represent experts, memory slots, or attention slots. Each row has unit probability mass, but no column receives positive mass from two rows; maintaining at least one vacant column makes assignments continuously deformable through elementary vacancy moves instead of abrupt softmax switches.
Construct a filtered cell complex from neural activations or a learned token/feature graph and track its persistence barcode incrementally as model activations change. Replace full persistent-homology recomputation at every checkpoint by maintaining homology bases and applying local transpositions when filtration blocks split or merge; use barcode drift as a training monitor or a weak regularization signal.
Add a measurement-conditioned attention layer with two explicitly separated fields: a geometry-only inverse-temperature profile that controls interaction strength and an outcome-dependent chemical-potential bias. For a region bounded by coordinates a and b, force the interaction gate to vanish as the square root of the distance from either boundary, while allowing a separate potential channel to encode measured values.
Use the paper's exponential dressing of an activity coupling as an adaptive gate on a neural network's nonlinear residual branch. The branch is strongly suppressed when the local activation fluctuation variance is high, producing an automatically linearized and more stable update, while low-variance representations preserve the learned nonlinear interaction.
Replace part of an attention matrix with a mixture of fuzzy permutation matrices induced by short permutations. Each basis element represents an order-preserving k-token matching smeared over all embeddings into the sequence, while a balancing constraint makes the aggregate attention receive uniform global coverage. Retain a standard low-rank or local-attention residual so the structured branch does not prevent arbitrary content-dependent interactions.
Replace a dense channel-mixing matrix in a sequence layer with alternating diagonal propagation and sparse unipotent Stokes jumps. The diagonal part carries independently controlled exponential phases, while the unipotent factors implement cheap residual-like mode conversion without changing determinant or requiring a dense matrix multiply. Constrain the phase magnitudes and jump coefficients during training to obtain a reversible, norm-monitorable mixer.
Add a small number of latent region-offset variables to a graph or token-mixing layer, interpreting selected edges as low-permeability barriers that suppress cross-region information flow. Eliminate the latent variables analytically, yielding a visible-node update with a structured low-rank correction rather than adding persistent hidden node states. The module is intended to preserve within-cluster propagation while preventing oversmoothing or contamination across learned boundaries.
Replace a softmax transition or mixture-of-experts router by probabilities obtained from squared amplitudes of an isometric latent transition. Each input state is mapped to an orthogonal latent subspace, and summing probability over the latent index produces the desired expert or next-state distribution. The latent amplitudes can retain information that would be destroyed by directly averaging expert outputs, while normalization is guaranteed by construction.
Build a parameter-free spectral channel mixer whose channels are arranged as components of an l-form and whose multiplier is the trace-free Beurling--Ahlfors transform. At every nonzero spatial frequency it mixes the exact and coexact channel subspaces with opposite signs, preventing a uniform channel-direction bias and preserving a structured cancellation property. Insert it as a residual branch before a convolution, MLP, or attention block, with one learned scalar gate controlling its…
Attach each token or graph node a discrete grade a in a finite group A, and modify attention value composition with a normalized group 2-cocycle rather than independent pairwise gates. The cocycle provides a globally consistent projective interaction rule, so composing three messages gives the same result under either parenthesization. This may improve relational reasoning while reducing the number of freely learned interaction parameters.
Replace or augment an exponential state-space memory branch with a causal convolution whose lag-j weight is exp(-lambda j) ell(j)/j. The 1/j boundary provides broad logarithmic memory, while lambda supplies an explicit finite memory scale and prevents uncontrolled accumulation from an untempered long-memory kernel.
Apply a low-degree polynomial feature lift to normalized hidden representations and penalize degeneracy of the covariance in that lifted space. This can detect collapse in nonlinear combinations of features even when the raw hidden covariance appears healthy.
Estimate localized directional Fourier correlations between intermediate activations and their loss residuals, then penalize anisotropic concentration. The method can discourage unstable feature directions and improve robustness without requiring a full microlocal distribution reconstruction.
Insert a Walsh-Hadamard transform before a quantized categorical or activation bottleneck and assign coordinate-dependent quantization precision using the attenuation spectrum of a quaternary symmetric noise model. Coordinates corresponding to tensor-product frequencies with many nonzero indices are attenuated by higher powers of \(\delta\), so their quantization can be made coarser with little effect on the reconstructed post-noise representation. This creates a structured, fast transform…
Build a neural feature-mixing layer from a truncated shift S and a diagonal phase operator T satisfying TS=qST, with |q|=1. The relation forces moving one position in the graded feature basis to multiply the phase operator by q, providing a compact inductive bias for periodic, phase-sensitive, or cyclic data.
Add a low-dimensional spectral regularizer to an encoder or transformer representation by estimating the first N nonconstant modes of its Gaussian-weighted diffusion operator. Penalize excessive reciprocal spectral mass and unequal low-frequency eigenvalues, using a Gaussian-ball reference calibrated to the representation's effective mass; this discourages latent directions from becoming weak, collapsed, or strongly anisotropic.
Add an integer-lattice feasibility layer after ordinary top-1 or top-2 MoE routing. The router first produces its usual expert assignments, then minimally changes a small number of low-confidence assignments so the batch count vector lies in a prescribed lattice or desired coset, eliminating persistent modular load imbalance that ordinary auxiliary losses may not detect.
Replace raw hyperbolic embedding-radius regularization with a dimension-aware effective-radius target. For embeddings concentrated near hyperbolic radius rho in an n-dimensional hyperbolic space, regulate s times log(sinh(rho) / sqrt(n)) rather than rho itself, and use the same quantity to calibrate distance-logit temperature. This should make hyperbolic metric-learning behavior more invariant when embedding dimension, curvature, or model scale changes.
Regularize the Gram spectrum of selected neural layers so that its low-order moments match the spectral moments generated by a truncated q-boson Jacobi operator. Unlike a simple Frobenius or spectral-norm penalty, this controls several parts of the singular-value distribution simultaneously and can discourage harmful spectral tails without forcing all singular values to be equal.
Insert a short gKdV-inspired spectral flow between neural blocks to regularize rough feature maps without using an isotropic low-pass filter. The module applies a Fourier dispersive phase and derivative-coupled polynomial residual updates, with an optional finite factorial dilation penalty to encourage analytic-looking features.
Represent each token or graph node by an anti-Hermitian matrix latent state and replace a standard residual transformation with a discretized Lie-algebra vortex flow. The commutator nonlinearities are equivariant under global unitary conjugation, so the block can learn interactions without selecting a basis and preserves the anti-Hermitian state space when initialized there.
Construct a nested sequence of representations in which each coarser representation is obtained from the previous one by a 1-Lipschitz projection. Train prediction heads at multiple scales so coarse predictions remain stable and approximately recoverable from the finer representation, enabling early exit, token pooling, and controlled multiresolution inference.