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 full PSD self-attention with a pivoted Cholesky/Nyström approximation whose landmarks are sampled from the unexplained diagonal mass. Tokens with large residual self-similarity are more likely to become landmarks, so the rank budget is spent on difficult regions rather than uniformly selected tokens.
Train a fixed-rank neural weight update Y=USV^T with a projector-splitting Runge–Kutta step instead of independently applying Adam or gradient descent to U, S, and V. The update evolves the full low-rank matrix using the neural gradient but performs QR-based factor updates, avoiding S^{-1} and remaining stable when adapter singular values collapse or cross zero. Use a common-base midpoint construction so every internal stage starts from the same U,V basis and remains rank r.
Replace dense cross-attention weights with a balanced transport plan whose nonzero query-key edges are maintained by a multiscale active-set procedure. Solve the coarse token-group problem first, lift its support to the fine token grid, add only edges indicated by local cost or marginal residuals, and warm-start the fine problem from the lifted plan. This should provide a principled sparse attention pattern rather than fixing a global top-k pattern before seeing the transport solution.
Replace dense attention between tokens on opposite sides of a one-dimensional boundary or segment split with a dyadic low-rank approximation of a Cauchy/Hankel distance kernel. Each distance-scale block uses O(log(1/\varepsilon)) features, and the number of active scales grows only logarithmically with context length after discarding a narrow boundary layer. This is especially suitable for a relative-position attention branch or state-space-like long-range branch, rather than arbitrary…
Compress a transformer KV cache by selecting actual past tokens whose key or hidden-state columns form a stable basis for all cached tokens. Instead of retaining tokens with the largest attention scores or leverage scores independently, compute rank-revealing pivoting of the leading right-singular-vector matrix and retain its pivot columns, then evaluate attention using the representatives plus an optional low-cost residual correction.
Replace dense query-key attention with an adaptive cross approximation constructed from selected query and key pivot tokens. At each rank, choose the pivot pair that removes large estimated residual energy, update the residual by a rank-1 cross correction, and stop when the residual estimate reaches a target tolerance. The resulting factorization computes approximate attention using a small number of landmark interactions while adapting to the actual token distribution.
Replace selected dense neural-network operators by low-rank factors whose rank is selected by a randomized residual test at a user-specified tolerance. Construct candidate bases in large blocks for efficient matrix operations, then prune the block to the smallest rank that passes the residual criterion instead of treating the block size as the final rank.
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.
Replace a dense graph-attention or token-mixing matrix by a resolvent-like interaction operator and truncate it to graph neighborhoods whose radius is selected from an estimated spectral gap. Unlike fixed-window sparse attention, the sparsity level is tied to a measurable stability parameter and has an explicit exponential tail criterion.
Replace a dense coordinate-kernel interaction among N points by an orthogonal samplet transform with a sparse detail-detail matrix and a small polynomial branch. Detail basis vectors have vanishing moments, so smooth low-frequency behavior is represented by a few polynomial coefficients while localized residual interactions become sparse in the transformed domain.
Replace a sparse graph layer's separate edge transformations with one joint low-rank factorization of all transformations entering each target node. For target node i, concatenate the neighbor matrices horizontally, project all neighbor features into a shared low-dimensional receiving basis, and reconstruct one output; retain the self transformation exactly. This can reduce edge-parameter storage and message-passing FLOPs when the incoming block row has rapidly decaying singular values.
Approximate a dense symmetric interaction matrix in a neural layer by \(\widehat A=C\widehat M C^{\top}\), but compute the small core \(\widehat M\) from a two-sided sketched least-squares fit rather than from the landmark principal submatrix. This preserves signed or indefinite directions and avoids exploding outputs caused by an almost-singular \(A(I,I)\).
Construct a sparse attention variant whose positional interaction kernel is piecewise polynomial rather than an arbitrary softmax score. Store key-value content in exact monomial moments within a spatial or learned-coordinate box hierarchy, then evaluate all queries from those moments without materializing the query-by-key matrix.
Replace a global attention truncation rule with a per-query local-tail budget. For each query token, retain nearby or high-priority keys until the estimated discarded interaction strength is below a target epsilon; this uses the paper's central distinction between local tail mass and the extensive norm of the discarded operator. The resulting attention pattern can allocate long-range computation only to tokens whose local tail is large.
Make Laplacian latency attention hardware-friendly by quantizing the exponential decay and row normalization to powers of two. Exponentials and divisions can then be implemented with lookup tables and integer shifts, enabling a low-bit attention kernel for neuromorphic and edge accelerators.
Construct a symmetric token affinity matrix and approximate only its spectrally outlying token-mixing modes using a small set of sampled landmark columns. The resulting low-rank operator replaces an \(O(n^2)\) dense mixer by two skinny matrix multiplications, while the paper's residual guarantee predicts that large-magnitude global interaction modes are preserved.
Add a fixed or weakly parameterized residual mixer whose interaction between sequence positions at distance \(r\) is proportional to \(1/(r\log^2 r)\). Instead of truncating the kernel at a short radius, represent its heavy tail with dyadic distance bands and compute each band using prefix sums or block pooling, giving every token access to arbitrarily distant context at roughly \(O(L\log L)\) cost.
Replace dense grid tokens or global spectral features with coefficients of compactly supported kernels centered on a nested hierarchy of spatial points. Encode an input field into coarse-to-fine coefficients, apply a neural map to those coefficients, and decode the predicted coefficients at arbitrary query locations; the contribution from each level provides an explicit multiscale output decomposition.
Use the sharp exponential tail bound to set a local clipping threshold from a desired exceedance probability. Instead of globally clipping activations at a fixed value or percentile, clip each local window at its minimum plus B log(e/delta), where delta is the tolerated fraction of clipped entries.