analiza aiqbench
This paper offers a rigorous theoretical dissection of single-layer Transformers, interpreting causal self-attention as a linear recurrence within dynamical systems theory. It identifies specific operational regimes where softmax convexity constraints cause oversmoothing in linear dynamics or enable adaptive delay embedding for nonlinear state reconstruction under partial observability. These findings provide essential clarity on the representational boundaries and failure modes of attention mechanisms when modeling complex temporal data, moving beyond empirical black-box usage toward principled understanding.
A Mechanistic Analysis of Transformers for Dynamical Systems
arXiv:2512.21113v2 Announce Type: replace Abstract: Transformers are increasingly adopted for modeling and forecasting time-series, yet their internal mechanisms remain poorly understood from a dynamical systems perspective. In contrast to classical autoregressive and state-space models, which benefit from well-established theoretical foundations, Transformer architectures are typically treated as black boxes. This gap becomes particularly relevant as attention-based models are considered for general-purpose or zero-shot forecasting across diverse dynamical regimes. In this work, we do not propose a new forecasting model, but instead investigate the representational capabilities and limitations of single-layer Transformers when applied to dynamical data. Building on a dynamical systems perspective, we interpret causal self-attention as a linear, history-dependent recurrence and analyze how it processes temporal information. Through a series of linear and nonlinear case studies, we identify distinct operational regimes. For linear systems, we show that in the single-head attention-only setting, the convexity constraint imposed by softmax attention restricts the class of autoregressive operators that can be represented, leading to oversmoothing when the target dynamics require mixed-sign lag coefficients. For nonlinear systems under partial observability, attention instead acts as an adaptive delay-embedding mechanism, enabling effective state reconstruction when sufficient temporal context and latent dimensionality are available. These results help bridge empirical observations with classical dynamical systems theory, providing insight into when and why Transformers succeed or fail as models of dynamical systems.
Abstract:Transformers are increasingly adopted for modeling and forecasting time-series, yet their internal mechanisms remain poorly understood from a dynamical systems perspective. In contrast to classical autoregressive and state-space models, which benefit from well-established theoretical foundations, Transformer architectures are typically treated as black boxes. This gap becomes particularly relevant as attention-based models are considered for general-purpose or zero-shot forecasting across diverse dynamical regimes. In this work, we do not propose a new forecasting model, but instead investigate the representational capabilities and limitations of single-layer Transformers when applied to dynamical data. Building on a dynamical systems perspective, we interpret causal self-attention as a linear, history-dependent recurrence and analyze how it processes temporal information. Through a series of linear and nonlinear case studies, we identify distinct operational regimes. For linear systems, we show that in the single-head attention-only setting, the convexity constraint imposed by softmax attention restricts the class of autoregressive operators that can be represented, leading to oversmoothing when the target dynamics require mixed-sign lag coefficients. For nonlinear systems under partial observability, attention instead acts as an adaptive delay-embedding mechanism, enabling effective state reconstruction when sufficient temporal context and latent dimensionality are available. These results help bridge empirical observations with classical dynamical systems theory, providing insight into when and why Transformers succeed or fail as models of dynamical systems.
Machine Learning (cs.LG); Computational Engineering, Finance, and Science (cs.CE)
arXiv:2512.21113 [cs.LG]
(or arXiv:2512.21113v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2512.21113 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 315 cuvinte
