The paper addresses a core challenge in modeling dynamical systems: latent factors are often assumed to play a single, fixed role, but in reality their influence can shift across observations and over time. The authors argue for a more flexible view, where each latent factor's structural role is identified jointly with its dynamics.

To make this concrete, they offer a group-meeting analogy. An idea may be voiced by only a subset of participants, carry information from earlier parts of the conversation, and then go on to shape what follows. This suggests that a latent factor's role is not static—it can be local to certain observations, inherited from the past, and influential on the future.

The abstract is truncated, so details of the method and experimental results are not available here. What is clear is the motivation: a more nuanced account of where latents belong and how they flow through a system.