Prune redundant attention heads using separate similarity scores for sink behavior and content routing. Two heads are considered safely redundant only when their normalized content compositions are close in Aitchison distance and their sink-mass trajectories are also close, avoiding pruning decisions dominated by a shared sink token.
Freeze a randomly initialized single-layer transformer and use a constructed soft prompt to make its attention weights equal Gaussian-kernel weights over support examples. The resulting model performs Nadaraya-Watson regression in one forward pass, so task adaptation stores prompt tokens rather than modifying network weights. Prompt length becomes the number of kernel centers, while hidden dimension and prompt norm determine whether the required logits can be represented accurately.
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.
Use attention-graph hitting times to identify tokens whose information has not mixed through the network, then route only those tokens through additional Transformer blocks. Tokens with fast reachability exit early, while slow or isolated tokens receive more computation.
Augment a causal sequence model with a small hierarchy of prefix summaries weighted by powers of the logarithmic rank of each preceding token. The summaries retain order-sensitive deviations from a baseline representation while costing O(KNd) for sequence length N, hierarchy width K, and hidden dimension d, instead of O(N^2d) dense attention.
Represent a nonnegative attention or routing score matrix by two nonnegative vectors, X = uv^T, and learn only entries on a sparse bipartite graph of important query-key or token-expert interactions. Complete the remaining entries multiplicatively and monitor cycle residuals as a certificate of whether the sparse representation is compatible with rank one. Use local ratio violations to trigger additional edges or relax the rank-one approximation only where needed.
Replace dense token-to-token attention with attention over connected token groups, called polymers, while forbidding nearby polymers from being simultaneously selected. Each candidate group receives an exponentially decaying size and boundary penalty, and the layer sums or samples only compatible collections of groups. The construction should create structured sparsity and prevent redundant overlapping attention regions.
Replace an unrestricted collection of nested dyadic attention windows on a 2D token grid by a sparse antichain: no selected window may contain another selected window. Use the paper's exponential occupancy guarantee to control how many attention blocks reuse the same token, and add a differentiable log-moment penalty during training when exact antichain selection is relaxed. The expected benefit is bounded peak KV reuse and more predictable sparse-attention cost without discarding multiscale…