KArAt proposes learnable attention without softmax

Published on October 27, 2025, the account presents Kolmogorov-Arnold Attention (KArAt), a learnable alternative to attention with fixed softmax. The proposal explores potentially more inspectable attention interactions and describes low-rank variants intended to reduce memory use.

Reported results in vision transformers are mixed, and the text notes trade-offs across model sizes; it does not identify a universal winner. To assess the claims, consult the original paper and post and check their experiments, configurations, and reported measures. If using AI to study or apply the method, do not submit personal data or confidential material without authorization.