Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
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 free-form order-dependent gate with a positive mixture of Mellin powers $(1+s)^{-a}$. This gives a small, interpretable module whose response across the order variable is automatically generated by a positive measure and therefore inherits complete monotonicity, log-convexity, and Hankel-moment structure.
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.
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.
Add a cheap spectral gate to a state-space model or recurrent event detector that decides whether multi-step lookahead can change the threshold decision. If the learned threshold readout is approximately a nonnegative left eigenvector of the transition matrix, use the current state only; otherwise activate predictive heads and search over a small horizon. This avoids unnecessary rollout computation while preserving early-warning behavior in oscillatory or rotating dynamics.
Treat a recurrent or state-space layer as a finite-state Markov cocycle and constrain optimizer steps using the paper's inverse-logarithmic sensitivity of Lyapunov exponents near a zero exponent gap. Instead of enforcing a crude spectral-norm bound, allow updates that are harmless for long-run growth while shrinking steps that could substantially change the recurrent stability profile.
Replace a one-step recurrent update with a causal convolution over past affine innovations using an exponential-fractional kernel. Add mean reversion and calibrate the innovation amplitude so that activation mean and variance remain approximately invariant across sequence position while retaining long-range, power-law-like memory.
Regularize hidden-state trajectories so that their temporal statistics match the moments of an isotropic Haar-distributed state up to order k, while deliberately leaving moments above k unconstrained. Use k as a controllable mixing knob: k=1 or 2 suppresses drift and anisotropic variance, whereas larger k imposes stronger distributional invariance and may remove useful temporal information.
Replace the fixed decay coefficient of a stochastic recurrent or state-space layer by an adaptive mean-reversion coefficient driven by the cumulative squared hidden-state energy. The controller approximates conditioning the latent trajectory on a small L2 norm: high-energy trajectories receive stronger restoring drift, whereas low-energy trajectories retain the base dynamics and noise.
Replace an unrestricted additive recurrent or fast-weight memory with a sign-selectable update: for each incoming update vector, choose between adding and subtracting it so that a smooth compact potential of the memory state is minimized. This is appropriate when the memory representation has sign symmetry, such as signed random features or a learned linear sketch; it is not a drop-in replacement for ordinary gradient updates where the sign carries semantic information.
Construct a branching residual network whose active computational paths reproduce according to a fixed offspring/connectivity law, while a controller can only remove paths using an age- or depth-dependent hazard \(u(a)\). Use the resulting bound as a diagnostic and gating schedule: removal can suppress unstable activity and reduce compute, but it should not be expected to cross the reproduction-driven propagation barrier unless the network's expansion operator is also changed.
Replace unconstrained recurrent-state decay with a one-dimensional latent defect field whose states evolve by local diffusion and pair reactions. Defects can move over long distances and persist, while creation and removal occur only in pairs, giving the memory a structured cancellation mechanism that is potentially better suited to delayed-event and parity-like sequence dependencies than a standard GRU or diagonal SSM.
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.
Replace a deterministic population activation or router fraction by a finite-population random rate whose noise is derived from an explicit binomial transition law. The layer preserves the desired mean activation while injecting variance that decreases with population size, creating a controllable stochastic bottleneck rather than uncalibrated Gaussian noise.
Replace a uniformly discretized recurrent or continuous-depth model with hybrid hidden-state dynamics: integrate a learned drift between event times, then apply a one-sided reflection update at each irregular observation or constraint event. The reflection prevents the hidden state from violating a lower obstacle, while the explicit jump decomposition avoids smearing abrupt information changes across many small residual steps.
Replace a single recurrent transition with K mode-specific neural transitions and train them using mode-aware normalization derived from the effective sample size T p_i. The model explicitly preserves the distinction between frequent and rare dynamical regimes, preventing frequent modes from dominating the shared training objective while avoiding unstable updates for poorly observed experts.
Construct augmented sequences by tiling long contexts with large finite blocks sampled from a library whose local-pattern frequencies match a target dataset, replacing only block interiors and leaving boundary zones untouched. This produces globally coherent synthetic contexts while controlling the distortion of short-range statistics through an explicit boundary-to-volume ratio.
Add a mean-field stochastic binary recurrent layer with an explicit susceptibility controller. The layer estimates the response statistic \(\chi=\beta^2N^{-1}\sum_i\operatorname{sech}^4(u_i)\) and either penalizes or clips it below \(1-\delta\), preventing the high-gain regime in which replicas with identical weights develop strongly divergent states. The expected benefit is more stable long-horizon recurrence and lower variance across stochastic forward passes.
Split a recurrent state into two blocks and initialize their variances and cross-correlation according to the stationary projective energy distribution induced by the transition. This places the initial hidden state near the typical invariant direction of the dynamics instead of forcing a long transient from zero or isotropic noise.