Localmax dynamics for attention in transformers and its asymptotic behavior
This work provides a theoretical framework for understanding attention mechanisms in transformers, but it is incremental as it builds on existing opinion dynamics methods.
The authors tackled the problem of modeling attention in transformers by introducing localmax dynamics, a discrete-time model that interpolates between softmax and hardmax dynamics, and they proved that its asymptotic behavior involves convex polytope convergence with quiescent sets, without finite-time convergence.
We introduce a new discrete-time attention model, termed the localmax dynamics, which interpolates between the classic softmax dynamics and the hardmax dynamics, where only the tokens that maximize the influence toward a given token have a positive weight. As in hardmax, uniform weights are determined by a parameter controlling neighbor influence, but the key extension lies in relaxing neighborhood interactions through an alignment-sensitivity parameter, which allows controlled deviations from pure hardmax behavior. As we prove, while the convex hull of the token states still converges to a convex polytope, its structure can no longer be fully described by a maximal alignment set, prompting the introduction of quiescent sets to capture the invariant behavior of tokens near vertices. We show that these sets play a key role in understanding the asymptotic behavior of the system, even under time-varying alignment sensitivity parameters. We further show that localmax dynamics does not exhibit finite-time convergence and provide results for vanishing, nonzero, time-varying alignment-sensitivity parameters, recovering the limiting behavior of hardmax as a by-product. Finally, we adapt Lyapunov-based methods from classical opinion dynamics, highlighting their limitations in the asymmetric setting of localmax interactions and outlining directions for future research.