✗ 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
✗ Mechanism failed
2026
Replace a first-order optimizer update by an extrapolation point followed by one damped Newton or Newton-CG solve, while selecting the acceleration weight from an explicit cubic Hessian-Lipschitz budget. Use a displacement-based safeguard in place of the unavailable distance to the optimum, turning the proof condition into a practical trust-region-like rule that limits unstable momentum.
Useful8/10
Difficulty6/10
Novelty6/10
✗ Failed on benchmark
2026
Replace the global EMA update for each linear-layer momentum matrix with a delta-rule update that learns the current output-side gradient value only along the current input-key direction. Frequently occurring directions are corrected repeatedly, while rarely visited directions are not unnecessarily overwritten or uniformly decayed. Use the resulting matrix as the ordinary momentum buffer in SGD, AdamW, or another optimizer.
Useful8/10
Difficulty5/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Replace a fixed or percentile-based gradient-clipping threshold with a threshold computed from the exact joint bias-energy envelope. The controller allows the user to specify how expensive removed-gradient bias is relative to retained update energy, while a running p-moment estimate determines the radius needed to satisfy a target joint-cost budget.
Useful8/10
Difficulty4/10
Novelty6/10
✗ Failed on benchmark
2026
Replace full-precision communication in decentralized or federated optimization with a sparsified uniform quantizer whose scale decreases geometrically, while maintaining an error state at each worker. Choose the scale so that quantization disturbance decays at least as fast as the contraction of the gradient-tracking dynamics; this should preserve linear convergence instead of creating the usual fixed-quantization error floor.
Useful8/10
Difficulty5/10
Novelty5/10
✗ Failed on benchmark
2026
Replace a standard softmax-gradient update for a probability vector with a two-stage KL Mirror-Prox update. The predictor evaluates the population-dependent cost at the current distribution, and the corrector evaluates it at the predicted distribution, reducing oscillation when routing or attention costs are coupled across tokens or samples.
Useful8/10
Difficulty5/10
Novelty5/10
✗ Failed on benchmark
2026
Replace a fixed learning rate for each layer or parameter block with a bounded gain selected by the one-step-ahead predictive loss. The sign of the product between the current gradient and the next gradient estimates whether the previous update moved downhill: aligned gradients increase the gain, while sign reversals decrease it. A mirror-descent update on a bounded interval prevents the runaway step sizes that can occur with exponential or unconstrained learning-rate parameterizations.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Replace constant-radius SAM by a clipped radius that equals the usual radius when the gradient is large but shrinks quickly enough near stationary points. This preserves SAM's sharpness-aware behavior during most training while removing the nonzero stationarity floor caused by a fixed perturbation.
Useful8/10
Difficulty4/10
Novelty5/10
△ Mechanism confirmed, baseline not beaten
2026
Replace Adam's elementwise denominator on factor matrices with a gauge-equivariant denominator shared across latent coordinates, or with a left preconditioner constructed from Gram matrices. The resulting update treats all equivalent bases of the latent space identically, allowing the low-rank implicit bias of factorized gradient flow to survive adaptive optimization.
Useful8/10
Difficulty4/10
Novelty7/10
✓✓ Beats tuned baseline
2026
Replace full-KKT implicit differentiation through a constrained quadratic-program layer with differentiation through only the equality constraints and inequalities active at the optimum. The forward solver still enforces all constraints, but the backward linear system scales with the active-set size rather than the total number of inequalities.
Useful8/10
Difficulty5/10
Novelty5/10
✗ Failed on benchmark
2026
Train the raw neural policy to anticipate the safety shield instead of relying on it indefinitely. Add the expected process-correction distance to the PPO objective and adapt its multiplier with a primal-dual update so that policy correction remains below a specified budget.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Compress the matrix gradient or momentum before applying Muon's polar LMO, and maintain an error residual in the uncompressed gradient space. The residual prevents systematic sign quantization bias from accumulating, unlike error feedback applied after the nonlinear polar/sign operation. This is suitable for distributed training because workers communicate one sign bit per matrix entry while the server still applies a matrix-aware Muon step.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Before quantizing a matrix product, reparameterize its factors as A'=AT and B'=T^{-1}B, preserving the exact full-precision product while changing the quantization difficulty of each factor. Choose a positive diagonal T=diag(t_1,...,t_K) that minimizes predicted post-quantization product error, rather than using output-channel scaling or a fixed heuristic grid. The gauge can be shared across several products when transformed-copy cost matters.
Useful8/10
Difficulty5/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Replace a fixed optimizer preconditioner with a diagonal matrix selected by an online convex optimizer. A gradient predictor supplies the direction, while a linear-loss regret update learns coordinate-wise gains that favor transformations aligned with the realized stochastic gradient. The method retains the identity preconditioner as an explicit comparator, so it can be tested for negative regret and improvement over ordinary SGD.
Useful8/10
Difficulty4/10
Novelty6/10
△ Mechanism confirmed, baseline not beaten
2026
Replace Euclidean or entrywise Kronecker fitting of a layer curvature matrix with its affine-invariant projection onto G = A tensor B. Use the resulting factors as a compact SPD preconditioner in the optimizer, while solving the projection through logarithmic residual partial traces and Armijo line search.
Useful8/10
Difficulty6/10
Novelty6/10
△ 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
✗ Mechanism failed
2026
Replace the ordinary average of task or client gradients with an iterative spectral filter that removes tasks whose gradient vectors explain an anomalously large covariance direction. The global model uses the filtered gradient, while each task still maintains its own personalized parameters and local optimizer state. Unlike parameter-center regularization, the robustification acts directly on the vector messages and is designed to avoid an additional \(\sqrt d\) contamination penalty.
Useful8/10
Difficulty5/10
Novelty5/10
✗ Mechanism failed
2026
Replace a full neural-network Gauss–Newton solve with a damped solve in an adaptively constructed low-dimensional parameter subspace. The subspace contains the current gradient, recent accepted updates, Krylov curvature directions, and randomized Jacobian-curvature probes, and is enlarged whenever its projected gradient fails to capture enough descent information.
Useful8/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Replace ordinary global gradient clipping with clipping of each stochastic gradient around a robust minibatch center rather than around zero. This preserves the common directional component of the gradients and suppresses only heavy-tailed residuals, making the update usable when gradient noise has a finite α-moment for 1 < α ≤ 2 but no finite variance.
Useful7/10
Difficulty4/10
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 failed
2026
Replace a fixed momentum and learning-rate schedule with a batch-aware stability controller derived from the paper's critical-learning-rate scalings. Polyak learning rates should scale approximately with B(1-rho), whereas Nesterov learning rates can scale as B^beta(1-rho) until reaching the base stability ceiling; this may allow larger batches without crossing the instability boundary.
Useful7/10
Difficulty5/10
Novelty6/10
✗ Mechanism failed
2026
Replace a fixed PAGE refresh schedule with a restart policy selected from the PL condition-number regime. For well-conditioned objectives, use frequent full-gradient refreshes and short inner phases; for ill-conditioned objectives, use the conventional condition-number-scaled PAGE phase length. The goal is lower component-gradient cost to a target loss while retaining PAGE's low-variance updates.
Useful7/10
Difficulty5/10
Novelty5/10
✗ Mechanism failed
2026
Replace point-estimate expert routing with a nominal allocation and a least-favourable allocation under uncertainty in expert quality. If both allocations agree, use that route confidently; if they disagree, profile or evaluate only the expert-input pairs responsible for the disagreement. The same mechanism can be used offline to assign workloads to LLMs or online to choose among heterogeneous experts under a latency or FLOP budget.
Useful7/10
Difficulty6/10
Novelty7/10
✗ Failed on benchmark
2026
Modify an L-BFGS curvature pair only when the observed secant curvature is negative. Replace the gradient-difference vector by the smallest Euclidean or inverse-metric correction that enforces positive curvature, then use the unmodified BFGS update and two-loop recursion. This avoids the computational and conditioning cost of adding a large isotropic damping term to the whole inverse-Hessian approximation.
Useful7/10
Difficulty5/10
Novelty7/10