Machine-learning ideas tagged Inference speedup in the Solves taxonomy of the Math2NN corpus.
Replace a dense token-mixing, MLP, or low-rank-adapter projection with a Bernoulli-signed sparse matrix normalized by the paper's predicted operator scale. Estimate the active representation dimension and use the effective scale to detect regimes in which extreme sparsity is likely to cause unstable amplification or dead rows.
Replace an ordinary group-L1 penalty on structured neural components with a powered ratio-of-norms penalty applied to their nonnegative importance magnitudes. The ratio encourages importance to concentrate on a small number of heads, channels, or experts while being less sensitive to arbitrary rescaling of the underlying weights. After training, components with small importance can be physically removed and the model can be fine-tuned.
Add a local Jacobian spectral regularizer and an initialization sweep to steer a looped transformer away from uncontrolled near-unit dynamics. The goal is to prevent examples from entering a fold-critical regime with very long relaxation times, or alternatively to deliberately target a controlled critical regime when adaptive test-time compute is useful.
Replace a single plug-in top-k router decision with a confidence correspondence containing every router parameter candidate and sparse expert assignment that remains compatible with calibration and the current input. Project this set onto a hierarchy of expert groups and return the finest group-level decision supported by all surviving explanations; otherwise coarsen the route or abstain. Active endpoint bracketing evaluates only candidates that could still change the projected routing report.
Replace explicit thresholding of a large structured routing-logit tensor with the paper's cubic sign iteration applied directly to its TT cores. The resulting approximate sign field can be converted into binary expert gates or binary activation masks without materializing all tensor entries, enabling routing over exponentially large Cartesian index spaces. Use TT rounding after each Hadamard product to prevent rank explosion, and compare the method against dense thresholding on smaller…
Rank candidate prompts, demonstrations, critiques, or system instructions by how many bits of reproduction cost they save for a specified artifact distribution. Replace raw prompt-token heuristics with a paired score that rewards both higher success probability and lower generation computation, then train or retrieve prompts maximizing this score under a token budget.
Represent a 3D neural field using high-order polynomial coefficients attached to an adaptively refined tetrahedral mesh, with a small MLP predicting residual corrections from local coordinates. Refine only tetrahedra whose prediction, rendering, or PDE residual is large, and use globally ordered vertices so neighboring tetrahedra share identical face and edge coefficients without hanging-node constraint solves.
Represent adaptive computation or MoE routing as a continuous latent trajectory that crosses convex mode walls, with a Minkowski norm defining computational speed. At a wall, choose the outgoing latent velocity by the same constrained variational rule as billiard reflection, and use the local path-length certificate to detect or prevent pathological accumulation of infinitely many routing events in finite depth.
Add an asynchronous binary refinement module in which each spatial unit or graph node may change its predicted label once if its current label disagrees with a weighted neighborhood field, after which it is permanently frozen. This prevents recurrent flip-flopping in iterative segmentation or denoising and should preserve large-scale structures while allowing a final interface-localized correction phase.
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.
Build a recurrent ReLU module that evaluates a binary refinement cascade using a fixed-dimensional state and shared cell weights. Replace hard binary digit selection with two overlapping circle coordinates; switch between their affine state updates only at points where the two candidate updates agree, so the switch is an exact continuous piecewise-linear ReLU operation rather than a multiplicative gate.
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.
Compress a dense neural-network weight tensor with mode-wise selected fibers and a small sampled core, then convert the result into a Tucker layer for efficient execution. Select indices greedily using reconstruction residuals rather than random sampling, and fine-tune the compressed model to recover task accuracy.
Replace a dense block of a large neural-network weight matrix with an adaptively constructed cross approximation, but prevent premature termination using residual checks on O(m+n) additional diagonal-like entries. Accept a rank only after the probe residual has remained below tolerance for several consecutive iterations; otherwise continue adding pivots.
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 an unconstrained relative-position attention-bias table by a learnable two-dimensional Fourier representation, or regularize the bias toward a small Fourier ratio. The ratio favors coefficient concentration without depending on the absolute scale of the bias, allowing the trained bias to be reconstructed from a small number of dominant frequencies.
Compress a trained graph-attention model's edge-dependent logits or gates in the line-graph Fourier domain. Smooth edge values become concentrated in low-frequency coefficients, allowing low-frequency coefficients to retain more precision while high-frequency residuals use fewer bits or are discarded.
Replace the linear state transition in a small recurrent or state-space module by a circulant matrix acting on a vector over a finite field. The hidden state then has only finitely many possible values and follows an exactly periodic orbit after at most \(q^n\) states, eliminating numerical drift on modular-counting and symbolic-memory tasks. A learned real-valued encoder and decoder can surround the discrete core, while the transition itself is fixed, searched, or trained with a…
Replace raw neural scores with randomized pairwise-error probabilities relative to a reference candidate distribution. Use a fixed PEP threshold to accept, abstain, or form a variable-size candidate list; exact uniformity under the reference law makes the threshold interpretable independently of the model's score scale and robust to ties.
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 a dense channel-mixing matrix in a sequence layer with alternating diagonal propagation and sparse unipotent Stokes jumps. The diagonal part carries independently controlled exponential phases, while the unipotent factors implement cheap residual-like mode conversion without changing determinant or requiring a dense matrix multiply. Constrain the phase magnitudes and jump coefficients during training to obtain a reversible, norm-monitorable mixer.
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.
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…
Replace an eigendecomposition-based spectral controller in a small recurrent or state-space transition layer with explicit polynomial projectors. Each hidden state is split into invariant modes, and each mode receives a separately constrained recurrent multiplier, enabling direct suppression of unstable modes or selective retention of long-memory modes using only matrix-polynomial evaluations.