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.
Use the number of tokens participating in an attention operation to adapt the projected query/key dimension according to the sharp Johnson–Lindenstrauss bound. Short attention windows receive a dimension limited by n-1, while longer windows use a logarithmic dimension, avoiding a fixed large head dimension when the current attention set contains too few points to require it.
Approximate a long hidden-state or key/value trajectory by a small sum of exponential modes, storing only the mode nodes and vector weights instead of every timestep. Recover the modes from a Hankel matrix through an annihilating polynomial, then reconstruct the trajectory or a compact recurrent state during inference.
Replace uniform or random landmark selection in low-rank attention with approximate greedy pivoted QR selection on the attention matrix. At each step, select the key whose residual attention column has the largest norm, orthogonalize it against previously selected columns, and stop when the residual energy falls below a target tolerance. Use the selected columns as an adaptive attention basis and reconstruct the full attention output through a small coefficient matrix.
Use the paper's effective-rank and eigengap-dependent covariance estimation rates to construct a confidence-aware low-rank bottleneck for transformer activations or key/value tensors. The bottleneck is enabled only when the top-p empirical eigenspace can be estimated more accurately than the desired compression error; otherwise the layer remains full-rank.
Convert attention scores into binary incidence set systems at one or several score thresholds, then group queries that induce the same sampled-key pattern. Compute the expensive key-value aggregation once per pattern and reuse it for all queries in the group. The method is exact for the thresholded routing component and approximates dense attention when queries have a small number of stable high-weight neighborhoods.
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.
Sample one batching aggressiveness parameter once per serving epoch and use it for all busy periods, instead of independently randomizing every batch. This globally coupled randomness preserves nested batching behavior while preventing a fixed traffic pattern from repeatedly aligning with a deterministic timeout.
Replace coordinatewise rounding of activation or embedding vectors with nearest-point quantization in a learned full-rank lattice. Learn an affine transform that makes the empirical activation region more isotropic, while regularizing the lattice covering density so it does not become inefficient as dimension grows.
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.
Use the one-edit Hamiltonian walk as an explicit state machine for counters in autoregressive models or world models. Instead of regenerating and re-embedding an entire numeric string after every increment, update only the digit that changes, or append the single leading digit at a block transition.