Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Parameterize a large linear layer as a sum of binary tensor products, W = Σ_l A_l ⊗ B_l, and regularize a factor-level upper bound on its top-k singular-value sums. The bound controls all Ky Fan norms of W while requiring SVDs only of the small factors, making it suitable for tensorized MLP or attention projections.
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.
Augment standard pairwise attention composition with a shared ternary correction m_3(a,b,c), so two different parenthesizations of a three-token or three-edge composition need not agree exactly but are forced to differ by a learned bounded correction. This gives the network a structured way to represent context-dependent compositionality without replacing efficient quadratic attention with unrestricted third-order attention.
Replace a collection of overlapping sliding-window features with approximately orthogonal incremental features: the length-m feature contains information not predictable from shorter consecutive windows. Use the paper's transition-operator Toeplitz precision matrix to decorrelate the resulting sequence of window features before attention, suppressing duplicated local evidence and improving conditioning.
Insert a differentiable layer that enforces multiple affine consistency constraints by running several short strings of relaxed projections and averaging their outputs. Change the strings and weights across training steps, but impose bounded string length, positive averaging weights, and an almost-cyclic coverage rule so every constraint is revisited regularly. This creates an architecture-level analogue of dynamic string-averaging rather than applying one fixed projection order.
Instrument selected neural-network operators with cheap stochastic perturbations and estimate how much their outputs change under finite-precision perturbations. Use the resulting per-operator score to identify unstable kernels and selectively switch them to FP32, compensated accumulation, or a stable reformulation instead of running the entire model at high precision.
Build a low-rank trilinear attention module in which query, key, and value factors are constrained to the unit sphere and refined through a few proximal alternating sweeps. The proximal terms suppress factor oscillation and make each sweep improve a well-defined tensor interaction objective, offering a stable alternative to unconstrained tensor-power iterations.
Construct sparse attention masks as regular or almost-regular hypergraphs: every query selects exactly w keys, while every key receives nearly equal incidence within each score or locality class. This preserves a fixed sparsity budget while removing systematic positional and head-level load imbalance that can cause hot keys, uneven memory traffic, and unstable gradients.
Replace one-shot softmax expert routing with a small number of synchronous routing rounds in which each token resamples an expert with probability proportional to that expert's current load raised to a power \(\alpha>1\). The resulting positive feedback rapidly creates a dominant routing basin, potentially reducing the number of active experts and communication groups at inference while retaining a controllable exploration phase through the initial round or a token-specific score factor.
Augment ordinary edge or line-integral features with a learned second-level area term, then train the model to make predictions invariant to path subdivision and stable under nearby path perturbations. The second-level feature captures ordered pair interactions that a sum of edge increments cannot represent, providing a compact alternative to explicitly processing every intermediate point in a trajectory or graph walk.
Give each query-token pair a positive adaptive edge weight that evolves by a multiplicative rule instead of relying only on instantaneous dot-product attention logits. Edges whose aggregate interaction is useful can grow, while overloaded or incompatible neighborhoods can shrink. Sparse initialization is preserved because an edge initialized at zero remains zero under the multiplicative dynamics.
Replace independent binary early-exit or token-pruning decisions with a monotone randomized survival process for each token or expert route. A token can lose survival mass at each layer but cannot become active again; the model is trained with a reflected obstacle-style penalty that activates when the predicted value of continuing computation is below the value of stopping plus the compute cost. Mean-field statistics are computed over currently surviving tokens, making routing less sensitive to…
Add a learned scalar ordering to a directed graph attention layer and retain only forward edges, producing a DAG attention mask without requiring a supplied topological order. Train the ordering with a differentiable surrogate for weighted surplus, and regularize it toward the paper's explicit half-weight-minus-l2 certificate. This supplies a principled alternative to random masking or unconstrained bidirectional graph attention when causal or hierarchical information flow is desirable.
Replace a conventional softmax router or fixed halting score with a scalar confidence state that evolves as a bounded martingale diffusion. The state starts at the network's prior confidence, receives evidence-dependent stochastic increments, and is absorbed at 0 or 1; absorption selects an MoE expert or halts additional transformer blocks. State-dependent volatility lets the model explore aggressively when uncertain and commit rapidly when confident, while the martingale constraint prevents…
Replace unconstrained MoE router logits with structured phase scores indexed by N-subsets of M ordered parameters. Each token is assigned to the dominant phase, while neighboring routing regions obey the Grassmannian rule that adjacent labels share N-1 indices, reducing arbitrary fragmented decision boundaries and encouraging smooth expert transitions.
Replace unconstrained pairwise token grouping with a tree whose edges carry independent merge or cut variables. The connected components of the retained edges define a valid partition at every forward pass, while learned edge gates control the amount of token aggregation. A coarse component-level computation can then replace part of dense attention.
Compute separate doubled-angle orientation order parameters for left and right image regions, then expose their sum and difference as symmetric and antisymmetric global features. This gives a network a low-dimensional inductive bias for global vertical structure versus left-right imbalance, while retaining magnitude channels that indicate when either readout is undefined because orientations cancel.
Replace a dense block of a large neural-network weight matrix with an adaptively constructed cross approximation, but prevent premature termination using residual checks on O(m+n) additional diagonal-like entries. Accept a rank only after the probe residual has remained below tolerance for several consecutive iterations; otherwise continue adding pivots.
Replace or augment a residual neural layer with a Fourier-domain scale-selective flow containing a learned second-order term and a fourth-order stabilizer. The block permits controlled low-frequency amplification, as required by the KS infrared mechanism, while damping high-frequency feature noise and preventing unbounded spectral growth.
Require the calibrated state posterior to remain unchanged when evidence is presented through information-equivalent prompt templates. Compare state distributions after semantic aggregation rather than raw token probabilities, and add a total-variation consistency penalty during calibration or fine-tuning.
Add a continuation-value correction and an explicit positive margin to routing or adaptive-computation decisions that consume a stateful resource. The margin prevents the degeneracy in which learned rent exactly cancels immediate surplus, making the network highly sensitive to noise, quantization, or small state-estimation errors.
Construct attention groups as hyperedges of a linear r-uniform hypergraph: every pair of tokens is allowed to share at most one group, while each token participates in approximately the same number of groups. Apply local attention inside each group and aggregate the outputs across groups. The construction inherits the paper's sharp capacity bound and prevents both redundant pair interactions and high-degree token hubs.
Replace a dense channel-mixing matrix by a sparse complex generalised weighing matrix W with exactly w nonzero entries in every row and column, then use U=W divided by square root of w as a norm-preserving mixer. Restricting to k=2 gives a real matrix with entries in {+1,-1}; k=4 supports signed phase rotations. The exact isometry should preserve signal and gradient norms while reducing channel-mixing cost.
Wrap a nominal gradient-based optimizer with a diagonal sign matrix that flips updates independently for parameter blocks, while a scheduler tests candidate sign configurations using short-horizon decrease of a Lyapunov-like training energy. The wrapper never changes the magnitude of the nominal update, and when the effective sign pattern is constant, it should recover the behavior of the correctly oriented nominal optimizer after a finite search period.