△ Mechanism confirmed, baseline not beaten
2026
Replace a portion of quadratic key-value attention or an external episodic table with a per-sample matrix fast memory updated by rank-one delta corrections. The memory directly learns a linear key-to-value map and can be carried across sequence segments, providing cheap online adaptation with constant state size per head.
Useful7/10
Difficulty5/10
Novelty5/10
✗ Failed on benchmark
2026
Replace deterministic LoRA importance scores with one-sided tests of whether each rank-one update has population contribution at least a user-selected threshold. Maintain empirical contribution samples during fine-tuning, estimate their uncertainty, and prune the components with the weakest statistical evidence while respecting the target rank budget. The method should avoid deleting components merely because their latest minibatch gradient was small.
Useful7/10
Difficulty5/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Replace star-shaped parameter synchronization with a rooted-tree primal-dual optimizer in which each worker owns a parameter block and communicates only with its parent and children. Dual updates performed at a node are explicitly redistributed as child correction messages, preventing stale-consensus errors caused by level-synchronous execution.
Useful7/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Insert a slow routing state and an intermediate hysteresis variable between a neural memory and its next-state selector. The hysteresis prevents small prediction fluctuations from repeatedly changing the active attractor, while the slower router learns transition probabilities independently of the attractor parameters.
Useful7/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Train a matrix-valued neural layer under an exact or near-exact Stiefel constraint while using an l1 or row-group sparsity penalty. During early training, use manifold proximal-gradient steps to identify a stable nonzero support; once the support stops changing, switch to Newton-CG steps restricted to the smooth intersection of the Stiefel tangent space and the fixed-support subspace. This can reduce the number of optimizer iterations needed to obtain sparse, well-conditioned projections.
Useful7/10
Difficulty6/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
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.
Useful7/10
Difficulty6/10
Novelty5/10
✗ Failed on benchmark
2026
Treat every low-rank basis refresh as a change of coordinates instead of assuming that old optimizer coordinates remain aligned with the new basis. Transport the first moment with the basis-overlap matrix, but collapse the second moment to a rotation-blind isotropic estimate rather than applying the same coordinate transformation to elementwise squared moments. This should eliminate second-moment staleness while preserving the memory savings of low-rank optimization.
Useful7/10
Difficulty4/10
Novelty7/10
✓✓ Beats tuned baseline
2026
Replace full-precision all-reduce parameter averaging in synchronous distributed training with the paper's compressed gradient-tracking recursion. Each worker maintains a model state, a gradient-tracker state, and two communication memories; only compressed differences from the memories are exchanged, while the tracker preserves the global-gradient increment despite compression.
Useful7/10
Difficulty6/10
Novelty5/10
✗ Mechanism failed
2026
Use the activation singular-value spectrum to allocate a fixed zeroth-order parameter budget across layers instead of assigning the same rank everywhere. Layers with a large discarded singular-value tail receive more coefficient directions, while spectrally compressible layers use smaller adapters, preserving the gradient-relevant subspace under a global memory and query budget.
Useful7/10
Difficulty4/10
Novelty6/10
✗ Mechanism failed
2026
Replace magnitude-based channel or expert pruning with a subset-selection objective that maximizes the weakest direction in the candidates' activation span. Relax the binary mask to continuous gates, optimize an entropic soft minimum eigenvalue, and round the gates to retain a fixed number of channels or experts. This should preserve diverse representations and reduce redundant feature directions.
Useful7/10
Difficulty5/10
Novelty7/10
Audited (legacy)
2026
Replace uniform embedding dimensions with a globally budgeted allocation based on the estimated spectral complexity of each categorical feature. Tables whose category representations have large leading singular-value energy receive more dimensions, while high-cardinality tables are penalized because each extra dimension consumes more parameters.
Useful7/10
Difficulty4/10
Novelty6/10
✗ Mechanism failed
2026
Replace an opaque MLP vector field with a stack of trainable symbolic primitives that can express linear terms, monomials, products, and related analytic operations. Apply an L1 penalty and prune small primitive coefficients after rollout training, yielding a compact dynamics module that is cheaper to evaluate and easier to inspect.
Useful7/10
Difficulty6/10
Novelty6/10
Unverified
2026
Replace dense aggregation over patches, points, or graph neighbors with a state-dependent positive quadrature rule. A latent encoder produces the current manifold coordinate z, a small controller chooses a subset of entities, and a convex redistribution step assigns nonnegative weights so that the subset preserves the full aggregation on a learned feature basis. The method is especially suitable for neural operators and vision attention where many spatial entities are redundant along the data…
Useful6/10
Difficulty6/10
Novelty6/10
Unverified
2026
Construct each hierarchy transition as a coarse-graining map together with a local decoder that recovers the retained local representation. This creates a reversible-information path for features needed by later computation while still allowing high-frequency or locally irrelevant components to be discarded.
Useful6/10
Difficulty6/10
Novelty5/10
Unverified
2026
Replace unconstrained token-to-expert routing with a nonnegative mixture of a small set of feasible routing or communication patterns. Enforce resource limits using the paper's one-sided positive-contribution bound, which gives a conservative certificate without enumerating all joint token activation scenarios. Expand the pattern set only when a separation procedure finds a routing pattern that improves the router objective while adding useful capacity information.
Useful6/10
Difficulty6/10
Novelty7/10
Unverified
2026
Add a recurrent associative matrix to each selected transformer layer so recent key-value relationships can be retrieved without retaining every past token or performing gradient updates. The matrix uses input-dependent retention and write gates, but retrieval is always performed from the pre-write state, preventing the current target from leaking into its own prediction. Frobenius-norm clipping makes the recurrent memory bounded and provides a direct stability control.
Useful6/10
Difficulty4/10
Novelty4/10
Unverified
2026
Replace the dense output of selected linear projections with a two-sided magnitude threshold that emits zero for small values but preserves signed large values. Learn one positive threshold per projection, or optionally one threshold per output channel, so the network discovers where sparse events can be removed while retaining outlier information.
Useful6/10
Difficulty4/10
Novelty5/10
Unverified
2026
Use one fewer point than the exact support threshold for a weighted multi-output linear head. The paper proves that the worst-case multiplicative loss penalty at this budget is only 1+1/(dm^2), giving a principled memory-saving option rather than an arbitrary subset-size heuristic.
Useful6/10
Difficulty4/10
Novelty4/10
Unverified
2026
Extend regularized extragradient with proximal operators so nonsmooth penalties such as group sparsity, nuclear norms, or parameter constraints are applied at both prediction and correction stages. This can produce sparse or low-rank adapters while retaining the look-ahead stabilization for the smooth inner residual.
Useful6/10
Difficulty5/10
Novelty7/10
Unverified
2026
Represent a block of candidate neural updates or adapter components by symmetric influence matrices and select one sign for each component so their aggregate spectral effect is small. This imports matrix discrepancy into low-rank adapters, expert aggregation, or structured quantization, where controlling the worst direction of interference may be more useful than minimizing entrywise error.
Useful6/10
Difficulty5/10
Novelty7/10
Unverified
2026
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.
Useful6/10
Difficulty4/10
Novelty6/10
Unverified
2026
Replace fixed-budget token or patch pruning with greedy selection that combines a teacher-derived relevance score and Gaussian-process mutual information. Select an item when it is both relevant and non-redundant, and stop when the largest remaining information gain falls below a calibrated threshold instead of retaining a fixed number of items.
Useful6/10
Difficulty5/10
Novelty5/10
Unverified
2026
Train a graph neural network with a learnable fractional graph mask constrained toward the tree-double-matching polytope. This gives the model a global connectivity constraint while penalizing attention patterns that concentrate too many incident edges on individual vertices.
Useful6/10
Difficulty5/10
Novelty6/10
Unverified
2026
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 \(A(I,I)\).
Useful6/10
Difficulty5/10
Novelty6/10