Beyond Linearity in Attention Projections: The Case for Nonlinear Queries

• URL: http://arxiv.org/abs/2603.13381v1
• Date: Wed, 11 Mar 2026 03:13:10 GMT
• Title: Beyond Linearity in Attention Projections: The Case for Nonlinear Queries
• Abstract summary: In decoder-only and encoder-only transformers, the Query projection $W_Q$ may be set to identity without noticeable performance deterioration.<n>We replace $W_Q in mathbbRd times d$ with a nonlinear residual of form $Q(X) = X + f_(X)$, where $f_$ is a bottleneck with $d2 + O(d)$ parameters.<n>Experiments on GPT-3 small style models show consistent improvement over the baseline, comfortably outperforming a model with 12.5% more non-embedding
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent algebraic analysis shows that in decoder-only and encoder-only transformers, the Query projection $W_Q$ may be set to identity without noticeable performance deterioration. This is possible because attention depends on $X$ only through the products $XW_Q, XW_K, XW_V$, allowing basis transformations to be absorbed by adjacent layers and propagated through the network. We replace $W_Q \in \mathbb{R}^{d \times d}$ with a nonlinear residual of the form $Q(X) = X + f_θ(X)$, where $f_θ$ is a bottleneck MLP with $d^2 + O(d)$ parameters. The identity term anchors the nonlinearity to a known-good prior. Experiments on GPT-3 small style models show consistent improvement over the baseline, comfortably outperforming a model with 12.5% more non-embedding parameters. These results motivate investigation at larger scales and across modalities.

Related papers

• Same Flow, Different Paths: Variance Reduction in Flow Matching [67.12978375116599]
  We study the choice of this path from an optimization perspective.<n>Our main finding is that the choice of path $g_t$ can fundamentally change the convergence rate of SGD.
  arXiv  Detail & Related papers  (2026-09-15T15:01:04Z)
• Learning the Intrinsic Dimensionality of Fermi-Pasta-Ulam-Tsingou Trajectories: A Nonlinear Approach using a Deep Autoencoder Model [0.0]
  We find that the trajectories lie on a nonlinear manifold of dimension $mast = 2$ embedded in a $64$-dimensional phase space.<n>This dimensionality increases to $mast = 3$ at $= 1.1$, coinciding with a symmetry breaking transition.
  arXiv  Detail & Related papers  (2026-01-27T12:59:29Z)
• Closed-form $\ell_r$ norm scaling with data for overparameterized linear regression and diagonal linear networks under $\ell_p$ bias [0.0]
  We give a unified, high-probability characterization for the scaling of the family of parameter norms.<n>We then study linear networks trained by descent.
  arXiv  Detail & Related papers  (2025-09-25T13:59:22Z)
• Symmetry-Breaking Descent for Invariant Cost Functionals [0.0]
  We study the problem of reducing a task cost functional $W : Hs(M) to mathbbR$, not assumed continuous or differentiable.<n>We show that symmetry-breaking deformations of the signal can reduce the cost.
  arXiv  Detail & Related papers  (2025-05-19T15:06:31Z)
• Transformers as Support Vector Machines [54.642793677472724]
  We establish a formal equivalence between the optimization geometry of self-attention and a hard-margin SVM problem. We characterize the implicit bias of 1-layer transformers optimized with gradient descent. We believe these findings inspire the interpretation of transformers as a hierarchy of SVMs that separates and selects optimal tokens.
  arXiv  Detail & Related papers  (2023-08-31T17:57:50Z)
• High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation [89.21686761957383]
  We study the first gradient descent step on the first-layer parameters $boldsymbolW$ in a two-layer network. Our results demonstrate that even one step can lead to a considerable advantage over random features.
  arXiv  Detail & Related papers  (2022-05-03T12:09:59Z)
• Approximate Function Evaluation via Multi-Armed Bandits [51.146684847667125]
  We study the problem of estimating the value of a known smooth function $f$ at an unknown point $boldsymbolmu in mathbbRn$, where each component $mu_i$ can be sampled via a noisy oracle. We design an instance-adaptive algorithm that learns to sample according to the importance of each coordinate, and with probability at least $1-delta$ returns an $epsilon$ accurate estimate of $f(boldsymbolmu)$.
  arXiv  Detail & Related papers  (2022-03-18T18:50:52Z)
• Statistical Query Lower Bounds for List-Decodable Linear Regression [55.06171096484622]
  We study the problem of list-decodable linear regression, where an adversary can corrupt a majority of the examples. Our main result is a Statistical Query (SQ) lower bound of $dmathrmpoly (1/alpha)$ for this problem.
  arXiv  Detail & Related papers  (2021-06-17T17:45:21Z)
• Variance-Aware Confidence Set: Variance-Dependent Bound for Linear Bandits and Horizon-Free Bound for Linear Mixture MDP [76.94328400919836]
  We show how to construct variance-aware confidence sets for linear bandits and linear mixture Decision Process (MDP) For linear bandits, we obtain an $widetildeO(mathrmpoly(d)sqrt1 + sum_i=1Ksigma_i2) regret bound, where $d is the feature dimension. For linear mixture MDP, we obtain an $widetildeO(mathrmpoly(d)sqrtK)$ regret bound, where
  arXiv  Detail & Related papers  (2021-01-29T18:57:52Z)
• Estimating Stochastic Linear Combination of Non-linear Regressions Efficiently and Scalably [23.372021234032363]
  We show that when the sub-sample sizes are large then the estimation errors will be sacrificed by too much. To the best of our knowledge, this is the first work that and guarantees for the lineartext+Stochasticity model.
  arXiv  Detail & Related papers  (2020-10-19T07:15:38Z)
• How isotropic kernels perform on simple invariants [0.5729426778193397]
  We investigate how the training curve of isotropic kernel methods depends on the symmetry of the task to be learned. We show that for large bandwidth, $beta = fracd-1+xi3d-3+xi$, where $xiin (0,2)$ is the exponent characterizing the stripe of the kernel at the origin.
  arXiv  Detail & Related papers  (2020-06-17T09:59:18Z)
• Learning nonlinear dynamical systems from a single trajectory [102.60042167341956]
  We introduce algorithms for learning nonlinear dynamical systems of the form $x_t+1=sigma(Thetastarx_t)+varepsilon_t$. We give an algorithm that recovers the weight matrix $Thetastar$ from a single trajectory with optimal sample complexity and linear running time.
  arXiv  Detail & Related papers  (2020-04-30T10:42:48Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.