Theoretical understanding of self-attention has advanced along several fronts. One prominent line of work studies self-attention layers that admit an energy function, with state updates driving that energy downward. This energy-based view has become a common tool for analyzing layer dynamics.

The new paper steps outside that framing. It offers an 'energy-agnostic' perspective, analyzing recurrent self-attention dynamics through Jacobians instead of assuming an energy landscape. The abstract does not spell out the full results, but the positioning is clear: the authors aim to provide an alternative foundation for studying these layers.

The contribution is therefore primarily theoretical. Where earlier work leans on energy minimization, this approach works directly from the Jacobian structure of the dynamics.