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.
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…
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.
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.
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.
Route tokens to experts using Laguerre cells defined by the minimum control energy needed to move a token embedding to each expert prototype, rather than by Euclidean distance or an unconstrained learned router logit. Per-expert dual weights deform the cells so that minibatch routing follows prescribed expert capacities, giving a geometrically interpretable alternative to auxiliary load-balancing losses.
Convert a sequential modular network into parallel execution layers by placing mutually commuting operators in the same layer. The resulting circuit preserves all noncommuting precedence constraints while exposing safe concurrency and fusion opportunities for inference or training.
Build a multi-expert or multi-task layer whose feature channels are divided into a globally shared subspace and expert-private subspaces. Matrix-weighted message passing couples experts only through selected feature directions, while the nullspace preserves specialization; the graph-cut condition provides a concrete test that the shared channels can propagate across all experts rather than becoming disconnected islands.
Use the paper's finite-width O(n^{-1/2}) Gaussian-process approximation bound as a width-budgeting rule rather than choosing every hidden dimension uniformly. Estimate an architecture-specific constant for each layer or attention contraction, then allocate width according to the smallest dimension satisfying its allowed distributional error. This should produce narrower models at comparable GP-like behavior, or permit the same parameter budget to be concentrated in the layers where finite-width…
Replace dense coarse-to-fine cross-attention at multiresolution interfaces with a sparse, nonnegative overlap operator whose weighted feature average is exactly conserved between the two resolutions. Use this operator as a low-order path and blend it with an unrestricted neural cross-attention path through a convex limiter that keeps features inside a box or simplex domain. The construction is especially suitable for adaptive token grids, hierarchical graph neural networks, neural operators…
Estimate an expensive fine-model trace or quadratic-form quantity using a telescoping sum over cheap-to-expensive neural approximations. Allocate many probes to cheap levels and only a few probes to the expensive level, exploiting strong correlation between adjacent levels to reduce variance at fixed compute. Candidate levels include truncated Transformer depth, reduced width, low-rank curvature, coarser graph resolution, or progressively tighter implicit-solver tolerances.
Replace a learned binary MoE gate with a hyperplane whose two sides contain prescribed proportions of several token populations simultaneously. In a low-dimensional routing projection, solve the cap-volume equations directly, producing deterministic per-population load control without an auxiliary load-balancing loss. Recursively applying the construction yields a balanced binary expert tree.
Add a deterministic complex-valued state-space bank whose mode detunings become progressively smaller with mode index, Delta_n=c n^{-p}, while input couplings decay as B_n=b n^{-kappa}. For slowly varying or constant forcing, the summed state follows the paper's subresonant response and grows like t^{1-alpha}, providing controllable power-law memory with only O(N) recurrent state updates. This should improve long-context retention compared with a same-size unconstrained RNN or uniformly spaced…
Use the activation singular-value spectrum to allocate a fixed zeroth-order parameter budget across layers instead of assigning the same rank everywhere. Layers with a large discarded singular-value tail receive more coefficient directions, while spectrally compressible layers use smaller adapters, preserving the gradient-relevant subspace under a global memory and query budget.
Remove a latent relay or hub token from an attention or graph layer and replace its two-hop influence by direct effective edges between retained tokens. The correction is a normalized rank-one update, so it can preserve hub-mediated communication while reducing the number of stored and processed states.
Construct a sparse attention layer by sampling backward token histories as a continuous-time branching process rather than allowing every query to attend to every key. Each active ancestor either dies or branches into a bounded number of candidate ancestors, with branching probability controlled by a small parameter. The branch-out penalty predicts exponentially small probability of long, highly branching histories, providing a direct knob for receptive-field size and attention FLOPs.
Replace an unconstrained attention or MoE routing matrix with a row-stochastic matrix that preserves a learned Gibbs-like distribution over groups. Anneal its temperature so that the router converges to a block-upper-triangular operator: mixing remains flexible within equal-energy groups, while cross-group traffic becomes directional rather than oscillating or collapsing. This creates a hierarchy-aware inductive bias and can reduce the number of active expert or attention connections.
Replace dense token attention with K asynchronous local nonlinear consensus updates on a sparse token graph. Each selected token is moved to the unique incident p-mean of its neighbors, producing a tunable message-passing operator that is averaging for p=2 and nonlinear for other p.
Propagate uncertain inputs or parameter perturbations through a neural network with zonotopes, and choose the reduction rule after each layer using short lookahead that predicts downstream certified-bound error. Candidate reducers can include largest-generator retention, box conversion, and norm-based merging; unlike a fixed policy, the chosen rule depends on the current generator geometry and remaining network depth. The goal is tighter robustness bounds at the same generator budget, or the…
Replace dense long-range token interactions with sparse message passing on a balanced hierarchical tree. Boundary token features are extended to internal tree nodes by weighted harmonic relaxation, and the resulting weighted normal derivative is used as a learned multiscale interaction or regularization signal.
Replace heuristic softmax load balancing with a convex entropic projection that computes expert assignment fractions satisfying explicit linear capacity and load constraints. The projection is bounded componentwise and keeps poorly identified routing decisions away from extreme 0/1 allocations, reducing expert collapse and overload under noisy or shifted batches.
Construct a sparse directed attention mask over ordered tokens by treating token pairs as edges labeled by time or position. Apply a recursive three-hop spanner construction so that every dense temporal route has a sparse route using at most three times as many hops. Run attention only on the selected edges and use three sparse message-passing layers, or a small learned hop-composition block, to approximate dense causal attention.