Beyond Linearity in Attention Projections: The Case for Nonlinear Queries

March 11, 2026 ยท Grace Period ยท ๐Ÿ› the ICLR 2026 GRaM workshop: https://openreview

โณ Grace Period
This paper is less than 90 days old. We give authors time to release their code before passing judgment.
Authors Marko Karbevski arXiv ID 2603.13381 Category cs.LG: Machine Learning Cross-listed cs.AI Citations 0 Venue the ICLR 2026 GRaM workshop: https://openreview
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 $Q(X) = X + f_ฮธ(X)$, where $f_ฮธ$ 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.
Community shame:
Not yet rated
Community Contributions

Found the code? Know the venue? Think something is wrong? Let us know!

๐Ÿ“œ Similar Papers

In the same crypt โ€” Machine Learning