Time-series models are increasingly conditioned on known future inputs, such as control signals or planned actions. In controlled systems, these inputs are not fixed but can be changed, compared, and optimized before execution. This raises a new requirement: a model must do more than predict the future; it must also reason about the consequences of alternative actions.
To address this, the paper introduces ActTime, an action-conditioned world modeling approach for time series. The framework is designed to support the comparison and optimization of actions before they are executed, rather than treating them as static conditions. The abstract notes that this capability is essential for controlled systems, where decisions can be evaluated and refined in advance.
While the full details of the method are not provided in the abstract, the motivation is clear: future-aware models need to move beyond passive prediction toward active decision support. ActTime appears to be a step in that direction, offering a way to model time series with actions as controllable inputs.