Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Construct overlapping attention windows from a block design instead of using one dense sequence-by-sequence attention matrix. Every token appears in exactly $r$ windows and every token pair co-occurs in exactly $\lambda$ windows, giving uniform coverage and avoiding the uneven connectivity of arbitrary sparse masks.
Replace greedy uncertainty sampling with a shallow Monte Carlo Tree Search that plans sequences of neural-network data acquisitions using a propagated uncertainty state. Each hypothetical query reduces uncertainty at nearby or correlated points, so later rewards automatically penalize redundant coverage and include labeling, simulation, or trajectory-transition costs.
Use the paper's localized truncation residual as an online certificate for whether the current polynomial lift is expressive enough. Start with a low-degree edge lift and activate additional degree blocks or a learned closure only when the residual exceeds a calibrated threshold, avoiding the cost and instability of always using a large polynomial dictionary.
Replace full-history backpropagation through time for an online recurrent or state-space neural network with a fixed-length batch protocol. An encoder maps the most recent input-output window to the latent state at the beginning of each batch, after which the learned dynamics are rolled forward and updated recursively from the new batch only. This should prevent state drift across long streams while retaining adaptation to changing dynamics.
Replace repeated iterations of an expensive high-dimensional update S with iterations of a lower-dimensional latent map T, then decode the resulting latent state with D. Train E, D, and T with explicit intertwining losses so that encoding a full update agrees with updating the latent state, and decoding a latent update agrees with applying the original update.
Use the recursive errors-in-variables subspace spectrum as a controller for the width of a latent SSM rather than fixing the state dimension in advance. Neurons or state channels are added when corrected covariance eigenvalues rise above the noise floor and pruned when they remain below it, producing a model-order-adaptive recurrent architecture for nonstationary streams.
Use small-gain diagnostics to jointly learn module normalization and a communication partition rather than imposing a fixed global spectral constraint. Clusters should be formed around high-gain feedback loops, because grouping weakly related modules cannot improve the certificate and only adds bookkeeping.
Replace a capacity-penalty-only MoE router with a nonnegative shadow price for each expert, capacity bucket, or hardware resource. Route each token using predicted utility minus the relevant price, while computing a decomposed optimistic objective that certifies how much utility remains above the feasible routed value.
Grow a neural network by appending a trainable block together with an analytically initialized inverse block, so the newly added depth is exactly the identity at insertion time. After insertion, untie and optimize the two blocks independently; this preserves the current function while providing additional trainable degrees of freedom. For architectures with one expensive mixing operation followed by cheap channelwise blocks, the same construction can increase depth without repeatedly paying for…
Use the KS ratio to decide how many message-passing layers to execute per graph or per node, rather than selecting a fixed depth. In the subcritical regime, stop once the predicted remaining effect is below a tolerance; in the supercritical regime, continue until the observed logit change becomes small or a larger budget is reached.
Replace one globally shared LoRA adapter with a small set of adapters whose task membership is chosen by residual redundancy. Tasks with strongly correlated validation residuals share an adapter, while tasks with weak or antagonistic residual dependence receive separate adapters. Recompute the partition periodically so the architecture follows the coupling that remains after training rather than correlations in the raw labels or initial gradients.
Insert a topology-controlled strategic communication layer into graph neural networks: each node maps a bounded latent scalar to either a clipped amplified signal or an interval-quantized message, with the amplification determined by how much influence the receiver exerts on the sender. Weakly influential communication channels should become aggressively quantized, while highly influential channels retain more resolution. This creates a principled variable-rate message-passing architecture…
Treat parameter-space curvature modes as RG momentum shells and use a smooth cutoff to construct a scale-dependent preconditioner rather than abruptly clipping eigenmodes. The optimizer should expose measurable crossovers between overdamped, KPZ-like, and nearly inviscid relaxation, allowing the learning rate and damping to change at empirically detected transitions instead of following a fixed schedule.
Replace a conventional feed-forward block with a sparse temporal graph whose hidden units are shared across many computation paths. Each arriving message updates a shared accumulator, applies a nonlinear response, and schedules delayed messages to downstream neurons; constructive or destructive interaction emerges when multiple paths visit the same unit.
Construct a neural activation bottleneck by projecting hidden states into a fixed covariance-eigenbasis and retaining only the d largest-magnitude coordinates per sample. For Gaussian, decorrelated activations, the paper proves that adaptive top-d selection in the PCA basis has no greater expected residual energy than adaptive top-d selection after any other orthogonal rotation. This provides a principled alternative to learning an unrestricted rotation before sparsification.
Combine the very-weak residual with step activations and one-bit weights, so the deployed PDE solver uses threshold and binary operations while training still optimizes a differentiable surrogate. The weak objective only needs values of the trial function and therefore does not require differentiating discontinuous activations with respect to spatial coordinates.
Replace scale-blind pooling or downsampling with a coarse-graining block that carries an explicit relevant scale variable \(\eta\) alongside the feature field. The block is constrained to represent features in the memory-retaining form \(h(\xi,\eta)=\eta^{\alpha}F(\xi/\eta^{\beta})\), allowing both feature amplitude and profile shape to depend on the scale inherited from the input or previous RG step.
Treat unresolved inference items as active threats and allocate a fixed budget of C module evaluations per round. Each evaluation has an item-dependent probability of completing the item, while the scheduler observes only completion or failure after the round. Use fair allocation when completion probabilities are unknown or nearly homogeneous, then switch to a marginal-success greedy policy as feedback estimates become reliable.
Replace an unconstrained recurrent or state-space transition with a finite quadrature of completely monotone memory modes. Couple the visible state and memory states as adjoint operators, so their cross terms cancel in the energy derivative and the layer is contractive even when visible-state damping is zero.
Construct a periodic neural integral layer whose fixed singular kernel behaves like |y|^{-s} near the origin, but whose samples on the uniform grid are replaced on a small symmetric stencil by SinCoTrap correction weights. The correction cancels low-order Taylor errors caused by sampling the singularity, while all nonlocal grid points remain unchanged. Increasing the correction order from p=0 to p=1 or p=2 should reduce discretization error without increasing global grid resolution.
Constrain each member of a wide recurrent or neural-ODE population to use the same time-dependent vector field whose spatial components generate a finite-dimensional Lie algebra. Store m fundamental trajectories and one fixed invariant label per node, then reconstruct every node state with the Lie-Scheffers superposition map instead of integrating all n states independently. The resulting layer has an exact md-dimensional dynamical core and should preserve the full network trajectory up to…
Replace dense self-attention by a sparse mask on a one-dimensional or ordered token geometry, retaining local neighbors and adding long-range edges with probability proportional to distance raised to \(-(1+\sigma)\). Use \(\sigma\approx0.8\text{--}0.85\) as the initial regime because the paper finds that this range supports delocalized, GOE-like connectivity despite sparse bonds. The resulting layer has linear or near-linear attention cost while maintaining long-range paths.
Build a positively homogeneous convex network by representing every intermediate unit as a compact polytope and composing units with Minkowski sums, convex-hull unions, and positive dilations. This gives an explicitly convex and monotone architecture whose geometric complexity can be controlled independently of the number of sampled linear pieces, potentially producing smaller ICNNs for structured convex functions.
Construct a weight-tied transformer loop in which the recurrent state receives a bounded diagonal carry plus a learned block increment, rather than applying a residual identity inside the learned increment. Parameterize the carry so every channel is strictly below one, allowing many recurrent iterations without the state explosion observed with an unconstrained carry.