✗ Mechanism failed
2026
Cluster recurrent modules or MoE experts by the geometry of their observed finite-horizon input-output behaviors rather than by parameter distance. Train one shared optimizer/controller or low-rank adapter per cluster while retaining module-specific parameters and routing. This should reduce control and optimizer overhead without merging modules whose temporal responses are dynamically incompatible.
Useful8/10
Difficulty5/10
Novelty8/10
✗ Mechanism failed
2026
Replace a fixed first-order parameter update by a finite-horizon controlled local model for each important curvature mode of the network. The optimizer computes the Hamiltonian flow and its Riccati feedback gain; if the chosen horizon approaches a conjugate point, it shortens the horizon or increases control cost before the gain becomes singular. This converts the paper's finite-time transition into a measurable trust-region and scheduling mechanism for neural training.
Useful8/10
Difficulty6/10
Novelty8/10
✗ Mechanism failed
2026
Treat local neural-network training as a driven linear system and periodically modulate the learning rate by a small sinusoid. Estimate the transfer function from this modulation to loss or gradient observables, fit its relaxation poles, and set the learning rate below the measured instability boundary.
Useful8/10
Difficulty5/10
Novelty7/10
✗ Mechanism failed
2026
Approximate the minibatch loss Hessian by a positive-semidefinite bulk curvature plus a small signed transverse correction, and treat only the correction with explicit negative-curvature steps. This imports the paper's observation that all unstable directions can be confined to a low-dimensional subspace, producing a curvature-aware optimizer whose step-size boundary is governed by a small matrix rather than the full Hessian.
Useful8/10
Difficulty5/10
Novelty5/10
✗ Failed on benchmark
2026
Add a controllable delay to the gradient force during optimization so that parameters follow a delayed-gradient dynamical system. Choose the delay below the stability boundary for ordinary training, and temporarily cross the boundary when the optimizer is trapped in a sharp or stagnant basin, causing stochastic fluctuations to be amplified out of the basin rather than waiting for a rare Arrhenius escape.
Useful8/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Introduce an effective learning-rate, gain, or regularization parameter that follows the commanded target with a finite implementation rate, and compensate for its predictable threshold-crossing lag. The scheduler estimates the network's current spectral instability boundary and commands the target parameter to cross that boundary early enough that the effective parameter crosses it at the desired time, avoiding overshoot caused by optimizer or hardware smoothing.
Useful8/10
Difficulty5/10
Novelty7/10
✗ Failed on benchmark
2026
Model one period of a cyclic optimizer or periodically modulated recurrent network as a discrete-time linear time-periodic system obtained by linearizing the update around its current trajectory. Estimate a periodic Lyapunov matrix sequence and scale the next learning-rate or modulation amplitude so that every phase contracts according to a certified energy decrease. This should prevent delayed divergence caused by resonance with the schedule, even when individual phase Jacobians are…
Useful8/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Learn a low-dimensional Koopman operator from successive states of an iterative neural system, such as debate agents, recurrent refinement blocks, or diffusion denoising trajectories. Use the magnitude of the subdominant eigenvalue to predict the remaining number of rounds required for disagreement to fall below a target tolerance, and stop computation when the predicted deadline is reached rather than using a fixed round budget.
Useful8/10
Difficulty5/10
Novelty7/10
✗ Failed on benchmark
2026
Replace a slow sequence of resolvent or contractive fixed-point updates by a blockwise averaged-reflection extrapolation. The method computes reflected iterates R^j y_0, averages them with equal weights, and uses the result as the next macro-iterate. Unlike unconstrained Anderson acceleration, this construction has a uniform residual guarantee for every maximal monotone operator.
Useful8/10
Difficulty4/10
Novelty6/10
✓✓ Beats tuned baseline
2026
Replace full-state prediction in a neural simulator or neural operator with prediction of a perturbation around a cheap structured background trajectory. Compute the background defect and known linearized or nonlinear corrections explicitly, and let the neural closure model only the remaining residual. Add a residual-magnitude gate so the learned closure is suppressed when the structured solver already explains the target dynamics.
Useful8/10
Difficulty5/10
Novelty6/10
✗ Failed on benchmark
2026
Replace each recurrent neural state with two asymmetrically coupled variables: a slow state x_i and a fast momentum or drive variable v_i. Each coordinate or block updates independently using its locally available, possibly stale input; the auxiliary variable supplies inertia that suppresses harmful update-order sensitivity and can accelerate traversal toward a retrieved state or denoised solution.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Replace the uniform or power-law convolution in a recurrent or state-space layer by a Gaussian q-binomial fractional kernel with learnable order alpha and deformation q. The parameter q controls a concrete memory-localization transition: q close to 1 gives classical fractional power-law memory, whereas q<1 produces exponentially localized memory and should reduce long-horizon gradient interference and truncation cost.
Useful8/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Replace recurrence or nearest-neighbour analogue lookup with a learned delay-coordinate observer that continuously corrects a latent state using the current observation. Constrain the observer's closed-loop Jacobian or linear state matrix to have spectral radius below one, so prediction error contracts geometrically and required burn-in grows logarithmically with target accuracy.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Replace uniform collocation for a fixed random-feature neural PDE solver with sampling from the leverage-score density of the operator-applied features. Whiten the retained residual feature space before solving for output coefficients, so the sampled least-squares matrix has an identity-like expected Gram rather than inheriting severe anisotropy from the differential operator. The same construction can be used for a linearized neural network by treating Jacobian features as the trial functions.
Useful8/10
Difficulty5/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Replace simultaneous descent-ascent on a bilinear adversarial subproblem by an implicit midpoint step. The update is a Cayley transform of the skew-symmetric game Jacobian, so it rotates rather than amplifies oscillatory modes and remains bounded for arbitrarily large positive step sizes in the exact bilinear case.
Useful8/10
Difficulty6/10
Novelty6/10
✗ Failed on benchmark
2026
Compress each hidden layer by retaining directions that are simultaneously reachable from the observed input distribution and observable at the network output. Unlike PCA or SVD, the retained subspace is weighted by downstream task sensitivity, so high-variance but output-irrelevant directions can be removed while low-variance predictive directions are preserved.
Useful8/10
Difficulty5/10
Novelty7/10
✗ Mechanism failed
2026
Use the behavior-subspace gap as a trust-region constraint when applying a shared update to multiple recurrent modules or experts. A proposed common update is accepted only when post-update behavior subspaces remain close to their leader and their graph subspaces remain sufficiently transverse, preventing one shared optimizer step from destabilizing dynamically different members.
Useful7/10
Difficulty6/10
Novelty9/10
✗ Mechanism failed
2026
Replace Euclidean or unprojected covariance averaging with a projected Bures-Wasserstein barycenter layer. Each unit-step barycenter update is followed by eigenvalue clipping into \([\alpha,\beta]\), preserving positive definiteness and preventing ill-conditioning without an additional eigendecomposition.
Useful7/10
Difficulty5/10
Novelty6/10
✗ Failed on benchmark
2026
Use the paper's certified-well geometry to turn continuous localization into a cheap grid proposal stage followed by fixed-step refinement. Threshold the projection-residual score on a coarse grid, then run a bandwidth-calibrated gradient map only from accepted points and merge converged duplicates. This avoids dense optimization from every possible coordinate and is suitable for neural slot or source heads that must return a variable number of continuous locations.
Useful7/10
Difficulty4/10
Novelty8/10
△ Mechanism confirmed, baseline not beaten
2026
When a model, dataset, or parameterization has an involutive symmetry, transform gradients and curvature into symmetry sectors before applying adaptive updates. The Hessian and optimizer can then be handled as independent even and odd blocks, preventing curvature from one sector from forcing an unnecessarily conservative learning rate in the other.
Useful7/10
Difficulty4/10
Novelty6/10
✗ Failed on benchmark
2026
Replace random Fourier or random NTK features by a weighted deterministic quadrature rule for the kernel's feature integral. The resulting feature map uses the same linear-model interface as random features but can approximate the leading Gram-matrix eigenvalues substantially better at the same feature count, improving conditioning and reducing the width required for a target kernel approximation.
Useful7/10
Difficulty4/10
Novelty5/10
✗ Failed on benchmark
2026
Replace an instantaneous largest-eigenvalue learning-rate ceiling with a delayed-instability monitor for a slowly ramped optimizer or network gain. When a dominant complex eigenpair crosses from negative to positive real part, permit a controlled post-crossing interval, but stop or roll back when the accumulated positive growth budget exceeds the perturbation/noise margin. This exploits slow-passage delay without allowing unbounded training instability.
Useful7/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Replace a full Hermitian curvature matrix, such as a Hessian or empirical Fisher matrix, by its block-diagonal version only when the paper's perturbation certificate predicts a small eigenvalue change. Use the certificate online to merge poorly separated blocks and retain independent preconditioners for well-separated blocks, yielding a controllable accuracy-memory tradeoff rather than a fixed block-diagonal approximation.
Useful7/10
Difficulty6/10
Novelty5/10
△ Mechanism confirmed, baseline not beaten
2026
For a complex-valued recurrent or state-space layer, construct a positive envelope by replacing each factor matrix with its entrywise modulus. The envelope provably upper-bounds every entry of the complex product and therefore gives a cheap conservative estimate of worst-case amplification, while a learned phase-cancellation term can exploit complex interference without allowing unstable growth.
Useful7/10
Difficulty4/10
Novelty6/10