Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Use local split-fusion rewrites as a structured alternative to globally recomputing token clusters. A model proposes a small number of neighboring tree edits per input, accepts only valid edits that reduce a learned energy, and retains the previous hierarchy across layers or decoding steps.
Represent sequence positions as vertices on a circle and use a maximal family of mutually non-crossing d-arcs to define the allowed attention interactions. Rotate the family by d positions, or use several phase-shifted families across successive layers, so each layer has only O(N) edges but repeated layers propagate information over long distances. The geometric compatibility rule replaces arbitrary local-window or hand-designed sparse masks with a structured family whose maximality gives broad…
Replace one-step greedy landmark selection in Nyström attention or kernel compression with a restricted pairwise-lookahead rule. The lookahead is motivated by the paper's explicit obstruction: a signed triangle can make individual column gains exhibit increasing rather than diminishing returns, so the best next column need not belong to the best pair.
Replace a dense or irregular binary interaction matrix in a value-aggregation layer with a signed sum of blocky masks. Each blocky mask groups a set of query rows with a disjoint set of key columns, allowing all queries in a group to reuse one summed value vector. This is most suitable for linear attention, graph message passing, or any layer where the interaction matrix is applied directly to values rather than passed through a row-wise softmax.
Learn or select sparse cyclic convolution or relative-attention offsets whose pairwise differences collide less often modulo the sequence length. The paper's Fourier fourth-power identity turns this combinatorial objective into an FFT-computable differentiable loss, enabling fixed-K sparse patterns with lower aliasing and interference than random offsets.
Replace independent top-k expert or attention-edge selection with a gate vector constrained by hierarchical path budgets modeled on the paper's extended-Dyck-path polytope. Ordinary interactions receive continuous nonnegative capacities, while a designated class of cross-group interactions receives binary or clipped-to-one gates, producing structured sparsity and preventing many correlated paths from consuming the same routing budget.
Represent many related sparse graph or attention patterns inside one fixed host connectivity pattern and activate each target instance with binary directional masks. The learned edge transformation and sparse-kernel layout are shared across instances, while the mask selects the target graph, enabling one compiled operator to process heterogeneous structures.
Construct a candidate feature for every edge pair or structured token pair, then retain a numerically independent subset under a feature-Jacobian matroid. The neural layer computes only the selected interactions, preserving directions that add new information rather than pruning solely by magnitude or attention score.
Construct a sparse attention mask from a fixed regular candidate graph and one scalar random label per token, retaining edge $(u,v)$ when $x_u+x_v\geq\tau$. Unlike independent random pruning, this produces correlated neighborhoods and a controllable distribution of token degrees, potentially giving some tokens broad receptive fields while retaining a fixed sparse budget.
Replace an unconstrained order-k weight tensor with a sum of components that are symmetric only within selected contiguous index blocks. This preserves interactions between blocks while tying parameters under within-block permutations, providing a tunable middle ground between a fully dense tensor and a fully symmetric tensor.
Construct a sparse attention or message-passing mask by sampling edges with preferential weights (d_u+alpha)(d_v+alpha), while keeping the edge count below the predicted connectivity threshold. This creates hub-like local communication patterns but prevents one giant component from forcing dense information mixing, reducing attention cost and potentially mitigating oversmoothing.
Generate temporal attention or convolution weights with the Graham–Knuth–Patashnik recurrence instead of learning every lag weight independently. For nonnegative recurrence parameters, the resulting lag sequence is strongly log-concave, so its normalized kernel is naturally unimodal and suppresses high-frequency sign-free oscillations without requiring a separate smoothness penalty. The six parameters can be learned per head, channel group, or layer, giving O(1) learned parameters for an…
Represent a sparse higher-order attention head by a 3-uniform hypergraph whose hyperedge $(v,x,y)$ allows anchor token $v$ to aggregate a pairwise interaction between tokens $x$ and $y$. During mask construction, greedily reject edges that would create a $4$-cycle in the link graph $L(v)$, so the same pair of source tokens cannot reach an anchor through multiple redundant pairings. This produces a diversity-constrained sparse attention pattern with an explicit, measurable collision bound.
Represent a population of N circular latent states using a three-parameter Möbius transformation applied to fixed uniform reference phases, rather than learning N unrelated angles. The resulting states remain on the circle by construction and can model concentrated or nearly uniform phase populations through a single concentration parameter.
Tie neural parameters across feature channels according to the Schur multiplicity pattern of the fermionic coinvariant representation. In the two-fermion case, use one learned parameter block for each Schur degree instead of independently parameterizing every ordered pair of fermionic channels, eliminating redundant copies while preserving the relevant GL2 channel symmetry.