Audited (legacy)
2026
Use the attention probability distribution over an ordered context to choose contiguous token groups whose pooled attention masses have entropy as close as possible to a prescribed upper budget R. Replace the corresponding key/value vectors by one weighted representative per group, preserving token order and reducing the KV-cache length from n to m. Unlike unconstrained token merging, the entropy constraint gives a direct control knob over how concentrated or diffuse the retained attention…
Useful6/10
Difficulty5/10
Novelty7/10
✗ Mechanism failed
2026
Insert a weighted negative-semidefinite fourth-order mixing operator into a residual or state-space layer. Instead of learning an unconstrained token-mixing matrix, parameterize its dissipative component as Q = -a W^{-1} B^T W B, ensuring that this component cannot increase the chosen weighted feature energy. Use a boundary-aware finite-difference matrix B along the sequence axis, optionally with learnable banded coefficients while preserving the factorization.
Useful6/10
Difficulty5/10
Novelty5/10
✗ Mechanism failed
2026
Replace ordinary masked mean pooling with a Fourier-compressed quadrature operator for arbitrary two-dimensional or three-dimensional domains. The geometry is preprocessed once into reusable grid weights, allowing every channel and every training example using the same domain to be pooled without boundary-area bias.
Useful6/10
Difficulty3/10
Novelty7/10
✓✓ Beats tuned baseline
2026
Estimate the temporal spectrum of each sequence channel using a locally private procedure, then apply a regularized inverse-square-root spectral filter before the sequence enters attention or an SSM. The filter removes predictable low-frequency or narrow-band redundancy while avoiding unstable amplification at frequencies where the private estimate is small.
Useful6/10
Difficulty5/10
Novelty7/10
✗ Mechanism failed
2026
Replace unconstrained combinations of several attention or adapter operations with a brace-style composition that inserts each operation into a distinct ordered interval of a base sequence. The resulting computation preserves the order of host and inserted operations and forbids crossing dependencies, producing hierarchical attention patterns with an explicit structural bias.
Useful6/10
Difficulty6/10
Novelty6/10
✗ Failed on benchmark
2026
Replace the usual hand-designed expert-load penalty with a heterogeneous survival penalty derived from a susceptibility distribution. Each expert receives an availability factor q_e=G(A_e), where A_e is its cumulative recent routing pressure and G_e is a learned or fixed mixture of exponentials; highly used experts are suppressed smoothly, while heterogeneous experts can have different resistance to pressure. The mixture produces adaptive curvature and long-tailed penalties that may reduce…
Useful6/10
Difficulty4/10
Novelty6/10
✗ Failed on benchmark
2025
Replace a conventional scalar activation by a geometrically indexed family of affine pieces whose slope changes with the logarithmic magnitude of the input. The same two endpoint parameters are reused across all scales, giving a compact, explicitly scale-aware activation that can represent different responses for exponentially separated activation magnitudes.
Useful6/10
Difficulty4/10
Novelty7/10
Unverified
2026
Replace an unconstrained Cartesian-product router over heterogeneous branches with a router whose joint expert or state assignments obey a finite-group conservation rule. Branch i emits a distribution over labels in its own subgroup H_i of a common finite abelian group G; only tuples whose group sum is zero are retained. This gives an exact, differentiable structural prior for modular arithmetic, multi-relational graphs, multi-view fusion, or any setting where latent labels compose by a…
Useful5/10
Difficulty5/10
Novelty7/10
Unverified
2026
Represent an original neural-network block and a proposed rewritten block as constrained optimization formulations over inputs and trainable parameters, then certify that the rewrite preserves feasibility and the ordering of losses over a bounded domain. This gives a compiler or pruning pipeline a formal reject/accept gate instead of relying only on numerical regression tests.
Useful5/10
Difficulty7/10
Novelty8/10
Unverified
2026
Replace a static top-k MoE capacity rule with a router whose expert allocation evolves through a finite-domain coverage process. Experts with larger current occupancy can either receive more future capacity, intentionally amplifying specialization, or receive less capacity by reversing the size dependence, allowing a controlled test of the paper's asymmetry-amplification mechanism.
Useful5/10
Difficulty6/10
Novelty6/10
Unverified
2026
Build a constrained autoregressive model whose initial logits are generated from a translation-invariant MPS associated with a local zero-mode construction. The MPS supplies a structured valid distribution before a Transformer residual is added, so the model starts on the constraint manifold instead of learning validity through a penalty.
Useful5/10
Difficulty6/10
Novelty6/10
Unverified
2026
Partition activations into dyadic magnitude bands and allocate sparse connectivity separately to heavy and diffuse coordinates. Protect high-magnitude coordinates with more reliable connections while using randomized flat connectivity for the many small coordinates, keeping the total number of nonzeros fixed.
Useful5/10
Difficulty6/10
Novelty7/10
Unverified
2026
Replace ordinary per-coordinate activation normalization or covariance whitening with a learned or frozen affine map based on a minimum-volume ellipsoid calibrated to the activation convex hull. Estimate the hull's Minkowski symmetry and use the bound \(\rho=\sqrt{d/\alpha}\) to set a principled clipping radius, where \(d\) is feature dimension and \(\alpha\) is the measured symmetry. This should reduce worst-case clipping of asymmetric activation distributions without requiring a much larger…
Useful5/10
Difficulty6/10
Novelty4/10
Unverified
2026
Replace or augment standard sinusoidal or RoPE position features with bracket-quadratic phases $e(-\theta n\lfloor\beta n\rfloor)$ generated by a Heisenberg nilmanifold orbit. Multiple irrational coefficients and output frequencies produce a cheap deterministic encoding whose empirical cross-position correlations should exhibit cancellation instead of the periodic aliasing of rational or finite-frequency encodings.
Useful5/10
Difficulty3/10
Novelty7/10
Unverified
2026
Construct positional or relative-position features as a nonnegative mixture of lattice cosine functions instead of independently signed sinusoidal features. The resulting bias is the Fourier transform of a positive discrete measure with explicitly bounded spectral support, while the mesh and degree can be initialized in the paper's dense-but-controlled frequency regime.
Useful5/10
Difficulty4/10
Novelty7/10
Unverified
2026
Replace or augment a scalar periodic positional coordinate with a normalized bank of odd Fourier harmonics, keeping every position on the same-radius sphere. The resulting representation has an explicit translation-invariant similarity kernel, allowing the frequency count and spectral weighting to control how sharply attention distinguishes nearby versus distant phases.
Useful5/10
Difficulty3/10
Novelty3/10
Unverified
2026
Restrict a fine-tuning adapter or output head to the subspace invariant under a prescribed monodromy, analogous to the paper's unbroken flavor lattice. This removes update directions intentionally changed by the domain-loop transformation, producing a parameter-efficient adapter with an explicit algebraic constraint.
Useful5/10
Difficulty4/10
Novelty8/10
Unverified
2026
Insert a fixed reversible lattice shear into a residual network so successive blocks follow a structured monodromy orbit rather than using unrelated learned transformations. Apply the transformation to a small learned subspace of hidden channels while leaving the remaining channels unchanged. This creates deterministic phase-dependent feature mixing with no additional trainable parameters.
Useful5/10
Difficulty5/10
Novelty8/10
Unverified
2026
Replace an unconstrained three-token interaction block by three distinct pair maps constructed from anticommuting channel generators. For every token triple, enforce equality of the two composition paths A12 B13 C23 and C23 B13 A12, while retaining different parameters for the three edges. This creates a globally consistent three-way interaction without collapsing to a single shared pair operator.
Useful5/10
Difficulty6/10
Novelty9/10
Unverified
2026
Regularize learned skew generators so that their iterated Lie brackets span many independent feature-mixing directions rather than collapsing to commuting or redundant matrices. This turns the paper's controllability family into a differentiable diversity objective for structured neural layers.
Useful5/10
Difficulty4/10
Novelty7/10
Unverified
2026
Regularize hidden activations or per-example gradients with a discrete version of the paper's Z_E^2 norm. Apply an E-norm to the largest fraction of coordinates and an L2 norm to the remaining tail, allowing the model to preserve a few large responses while discouraging widespread heavy-tailed noise.
Useful5/10
Difficulty3/10
Novelty7/10
Unverified
2026
Add an auxiliary loss that makes selected representation coordinates insensitive to all subsets of fewer than d variables while retaining a d-way parity statistic. The objective discourages the network from solving a task through pairwise shortcuts and explicitly rewards a controlled high-order interaction.
Useful5/10
Difficulty5/10
Novelty7/10
Unverified
2026
Replace the ordinary triangle-inequality budget for merging m linear residual branches or LoRA updates by the sharp quasi-reverse Minkowski certificate. During training, penalize or constrain the Schatten norm of the aggregate absolute update, which certifies the norm of the actually merged update with factor C_{p,m} rather than the loose factor m. This is especially attractive for p=2, where the certificate controls Frobenius energy and can be implemented with standard matrix operations.
Useful5/10
Difficulty6/10
Novelty7/10
Unverified
2026
Insert a positivity-preserving fractional Schrödinger resolvent into a 1D neural sequence block. Given a nonnegative learned potential V, the layer transforms an input signal f using V^a(-Delta+V)^(-a)f, allowing the network to learn where to smooth or suppress features while retaining an L1 bound independent of the potential magnitude. Use a in (0,1] as a fixed hyperparameter or a clipped learned scalar.
Useful5/10
Difficulty6/10
Novelty7/10